Student Performance Prediction and Teacher

Recommender System

By

Muthara-Tul-Ain

MASTER OF SCIENCE IN COMPUTER SCIENCE

DEPARTMENT OF COMPUTER SCIENCE

CAPITAL UNIVERSITY OF SCIENCE AND
TECHNOLOGY
ISLAMABAD
2017
 Student Performance Prediction and
Teacher Recommender System

By

Muthara-Tul-Ain

A research thesis submitted to the Department of Computer

Science, Capital University of Science and Technology, Islamabad
in partial fulfillment of the requirements for the degree of

MASTER OF SCIENCE IN COMPUTER SCIENCE

DEPARTMENT OF COMPUTER SCIENCE

CAPITAL UNIVERSITY OF SCIENCE AND
TECHNOLOGY
ISLAMABAD
2017

i
 Copyright 2017 by CUST Student

All rights reserved. Reproduction in whole or in part in any form requires the
prior written permission of Muthara-Tul-Ain (MS133033) or designated
representative

ii
 CAPITAL UNIVERSITY OF SCIENCE & TECHNOLOGY
ISLAMABAD
Islamabad Expressway, Kahuta Road, Zone-V, Islamabad
Phone: +92 51 111 555 666, Fax: 92 51 4486705
Email: [email protected], Website: http”//www.cust.edu.pk

CERTIFICATE OF APPROVAL

Student Performance Prediction and Teacher Recommender

System
by

Muthara-Tul-Ain

MS133033

THESIS EXAMINING COMMITTEE

S No Examiner Name Organization

(a) External Examiner Dr. Ehsan Munir CIIT, Wah Cantt

(b) Internal Examiner Dr. Muhammad Arshad CUST, Islamabad

Islam

(c) Supervisor Dr. Nayyer Masood CUST, Islamabad

________________________________

Dr. Nayyer Masood

Thesis Supervisor
November, 2017
______________________________ ___________________________
Dr. Nayyer Masood Dr. Muhammad Abdul Qadir
Head Dean
Department of Computer Science Faculty of Computing
Dated : November, 2017
Dated : November, 2017

iii
 DEDICATED
TO

MY

RESPECTED

ALLAH (SWT) & PROPHET MUHAMMAD (PBUH)

&

PARENTS, BROTHER, SISTER

FOR

THEIR CONSTANT

AFFECTION AND SUPPORT

iv
ACKNOWLEDGMENT

All praise and exaltation is due to ALLAH (S.W.T) The creator and sustainer
of all seen and unseen worlds. First and foremost I would like to express my
gratitude and thanks giving to Him for providing me the boundaries and
blessings to complete this work. Secondly, I would like to express my sincerest
appreciation to my supervisor Dr Nayyar Masood for his directions,
assistance, and guidance. I sincerely thanked for his support, encouragement
and technical advice in the research area. I am heartily thankful to him from the
final level, as he enabled me to develop an understanding of the subject. He has
taught me, both consciously and unconsciously, how good experimental work
is carried out. Sir you will always be remembered in my prayers.

I am highly indebted to my parents and my family, for their expectations,

assistance, support and encouragement throughout the completion of this
Master of Science degree. They form the most important part of my life. After
ALLAH (S.W.T) they are the sole source of my being in this world. No words
can ever be sufficient for the gratitude I have for my parents and for my family.
A special thanks to my employing company for their support and
encouragement to complete this Master of Science degree.

I pray to ALLAH (S.W.T) that may He bestow me with true success in all
fields in both worlds and shower His blessed knowledge upon me for the
betterment of all Muslims and whole Mankind.

AAMEEN

Muthara-Tul-Ain

v
DECLARATION
It is declared that this is an original piece of my own work, except where
otherwise acknowledged in text and references. This work has not been
submitted in any form for another degree or diploma at any university or other
institution for tertiary education and shall not be submitted by me in future for
obtaining any degree from this or any other University or Institution.

Muthara-Tul-Ain

Reg. No, MS133033

November, 2017

vi
 ABSTRACT
Mining data and extracting information from huge databases has become an
interesting research area for the researchers. The idea to extract information
with the help of data mining techniques came into being since a couple of
decades ago. Initially, researchers were supposed to apply classification and
clustering techniques to partite the dataset and analyze the intrinsic features. On
the basis of such features, they make reasonable predictions. These predictions
have taken place in the field of educational data mining for many purposes such
as; predict the performance of students on the basis of factors associated with
them, to enable them suitable courses and appropriate teachers. These purposes
have been derived from the area of student retention and attrition. Our research
aims to achieve these purposes under the roof of student attrition and retention.
Moreover, we have identified such exiting factors which are beneficial for
predicting the performance to students, recommend them best suitable teachers
and help them to select the courses. We have applied classification algorithms
with respect to the nature of data which have been collected from the Capital
University of Science and Technology (CUST), ISB. In this study, the GPA of
first semester on the basis of Midterm and previous academic grades have been
tried to predict. For the second semester, we have predicted the CGPA of same
students by using their complete proceeding academic record with the help of
hybrid approach. The hybrid approach consists of combination of factors that
have been evaluated against our research questions. Moreover, we tried to
improve their performance by recommending those suitable courses and
teachers whose performance is better amongst others comparatively. The
reason of collecting the data from CUST is to validate the exiting factors in
local context (Pakistan). On the basis of classification algorithms such as;
Naïve Bayes and J48, we have become able to build the recommender system.
Then the factors which contributed well to validate the exiting factors in local
context have been measured. In the last, appropriate teacher allocation has been
measured by two ways; Statistical and Prediction. In statistical
experimentations, average performance of teachers, Z- test and ANOVA test
has been applied. In prediction experimentations, one subject teacher with other

vii
subject teacher’s name attribute, and one subject teacher without other subject
teacher’s name attribute and overall performance of teachers have been
computed with respect to each subject independently. With the help of our
research we might have become able to provide a way to educational
institutions to reduce the attrition and increase the retention rate.

viii
 Table of Contents
Abstract .........................................................................................................................vii
Introduction ..................................................................................................................... 1
1.1 Background Of Research ......................................................................................... 2
1.2 Problem Statement ..................................................................................................... 6
1.3 Research Questions.................................................................................................... 6
1.4 Scope ......................................................................................................................... 6
1.5 Application Of Proposed Approach............................................................................ 6
1.6 Significance Of Solution ............................................................................................ 7
1.7 Organization Of Thesis .............................................................................................. 8
1.8 Definitions And Frequently Used Terms .................................................................... 8
Literature Review ............................................................................................................ 9
2.1 Educational Data Mining ......................................................................................... 10
2.2 Student Performance Prediction ............................................................................... 14
2.3 Commonly Used Approaches In Edm ...................................................................... 16
2.5 Commonly Used Attributes For Student Performance Prediction. ............................ 19
2.5 Literature Review Summary .................................................................................... 21
Research Methodology .................................................................................................. 22
3.1 Data Collection And Pre-Processing ........................................................................ 24
3.2 Classification ........................................................................................................... 27
3.3 Algorithms And Techniques .................................................................................... 29
3.4 Evaluation Of Research Questions ........................................................................... 30
Results & Evaluations.................................................................................................... 35
4.2 Final Attribute Selection .......................................................................................... 38
4.3 Result Of Evaluation Of Research Questions ........................................................... 38
Conclusion And Future Work ........................................................................................ 67
5.1 Future Work ............................................................................................................ 70
References ..................................................................................................................... 71

ix
 List of Tables
Table 2.1: Literature Review Summary ........................................................................................... 21
Table 3.2: Summary Of Session-Wise Attrition............................................................................. 24
Table 3.3: Summary Of Session-Wise Attrition............................................................................. 25
Table 3.4: Selection Of Attributes ..................................................................................................... 28
Table 3.5: Attributes Selection To Dataset ...................................................................................... 31
Table 3.6: Pattern Of Combination Of Attributes .......................................................................... 32
Table 4.1: Occurrence Of Missing Values ....................................................................................... 36
Table 4.2: Handling Of Missing Values ........................................................................................... 37
Table 4.3: Attributes Table ................................................................................................................... 39
Table 4.4: Classification Of Attributes And Algorithms ............................................................. 40
Table 4.5 Ranked Attributes ................................................................................................................. 42
Table 4.6 GPA PREDICTION FOR THE FIRST SEMESTER .............................................................................. 42
Table 4.7: Cgpa Prediction For The Second Semester ................................................................. 44
Table 4.8: Pre-Qualification Vs Maths Courses ............................................................................. 47
Table 4.9: Average Of Cal-Ll Teachers ............................................................................................ 48
Table 4.10: Average Of Cal-I Teachers ............................................................................................ 49
Table 4.11: Overall Average Of Maths Teachers ......................................................................... 50
Table 4.12: Anova Analysis On Cal-1 .............................................................................................. 52
Table 4.13: Two Samples For Mean With Respect To Teacher “E”........................................ 53
Table 4.14: Two Samples For Means With Respect To Teacher “B” ..................................... 54
Table 4.15: Teacher B With Same Number Of Students ............................................................. 55
Table 4.16: Teacher B With Different Number Of Students ...................................................... 55
Table 4.17: Comparison Of All Programming Courses ............................................................... 56
Table 4.18: Comparison Of Actual And Predicted Grades Of Itc/Itp ...................................... 57
Table 4.19: Comparison Of Actual And Predicted Performance Of Oop/Cp
Teachers ............................................................................................................................................ 58
Table 4.20: Overall Performance Of Oop Teachers ...................................................................... 59

x
 List of Figures
Figure 3.1: DATA FLOW DIAGRAM OF PROPOSED METHODOLOGY ......................................................... 23
Figure 4.1: Formats Of Data................................................................................................................. 36
Figure 4.2comparison Of Local And Existing Attributes ............................................................ 41
Figure 4.3: Gpa Prediction For The 1st Semester .......................................................................... 44
Figure 4.4: Cgpa Prediction For The Second Semester................................................................ 45
Figure 4.5: Pre-Qualification Vs Maths Courses ........................................................................... 48
Figure 4.6: Average Of Cal-Ll Teachers .......................................................................................... 49
Figure 4.7: Overall Average Of Maths Teachers ........................................................................... 50
Figure 4.8: Comparison Of Itc/Itp Teacher ...................................................................................... 51
Figure 4.9: Comparison Of Actual And Predicted Grades Of Itc/Itp ....................................... 56
Figure 4.10: Comparison Of Actual And Predicted Performance Of Oop/Cp
Teachers ............................................................................................................................................ 58
Figure 4.11: Overall Performance Of Oop Teachers .................................................................... 59
Figure 4.12: Overall Performance Of Oop Teachers .................................................................... 60

xi
 List of Abbreviations
CUST: Capital University of Science and Technology

WEKA: Waikato Environment for Knowledge Analysis

MOOCs: Massive Open Online Courses

CSV: Comma-separated values (Comma-delimited)

ARFF: Attribute-Relation File Format.

GPA: Grade point average

CGPA: Cumulative Grade Point Average

Cal-I: Calculus-I

Cal-II: Calculus-II

ITC: Information Technology Computing

ITP: Information Technology Programming

CP: Computer Programming

OOP: Object Oriented Programming

SEM: Semester

xii
Chapter 1

INTRODUCTION
Researchers have been working in the area of educational data mining for a
decade. Educational Data Mining (EDM) has become a broadened field in
which researchers are conducting their experiments to extract useful
information from the data belonging to the educational sectors for many
purposes. The purposes include identifying student attrition and retention rate,
students’ performance prediction, building a course recommender system, and
teacher recommender system, etc. In the field of Educational Data Mining
(EDM), researchers are busy in exploring effectiveness or role of different
types of variables to measure and predict the performance of students. Among
those variables, student’s age, academic record and biography of students are
involved. EDM is a growing research field, which carries data mining
techniques in educational system (Romero, C., & Ventura, S. 2007). In
previous study, (Romero, C., & Ventura, S. 2013) has explored the
phenomenon of making the student’s outcome better with the help of data
mining approaches. According to them, huge data from the institutions have
problems associated with it. Therefore, data mining approaches are not
supposed to be applied directly on this data. Therefore, knowledge discovery
process has been implemented. In educational institutions, data mining is
playing a pivotal role on datasets after preprocessing. Most widely used data
mining approaches involved classification and clustering, outlier detection,
association rule mining, pattern mining and text mining. Researchers are trying
to extract useful, novel and interesting information with the help of data mining
approaches (Romero, C., & Ventura, S. 2010). EDM process is used to convert
raw educational data into useful information.

In the field of EDM, researchers are active in three broad categories; Personal
recommender system and learning environment, course management system,
and student attrition and retention. Our research will cover all these mentioned
areas in different contexts such as identification of students with low academic
performance, building prediction model to predict the student’s performance by

1
using their historical data, improvement in raising the confidence level among
them, assist them to choose their courses, building a model that will assign
appropriate teachers to the students.

1.1 Background of research

As discussed in above section, EDM is subfield of data mining which is applied
in educational institutions to help them to maintain their prestige. Educational
data mining is all related to propose and develop naval approaches to improve
the performance of students on a large scale. Educational data mining is an
active research areas in which researchers are working actively in three main
areas of educational data mining are:

a) Personal recommender system and learning environment

Researchers are using this concept in hybrid fashion in the context of
educational data mining. They are constantly using various data mining
approaches to make the personal recommender system better to produce
maximum accurate results. In a research (Miller, B. 2004), personal
recommender system uses collaborative filtering to reduce noise from the data.
A limitation of personal recommender system is found to be its un-portability
as they are operational only on large computers that are connected to the
internet. Another limitation was found to be its trustworthy relationship
between the owner of recommender system and its user.

b) Course management system and educational data mining

Researchers are associated with this branch of EDM to provide the universities
and educational institutions a platform that can help them in management of
courses in the department. For this purpose, different data mining approaches
have been carried out (Peña-Ayala, A. 2014). The concept has been derived
from e-learning system “Moodle”. Moodle is a learning management system
that helps educators to build up effective online learning communications
through the analysis of information generated by the users (Aher, S. B, 2012).

2
c) Student attrition and retention
Student attrition means the number of students who are leave their courses
without completing because of certain reasons. Those reasons may include poor
choices of courses, consistent undesirable result, poor performance, financially
instability and immature selection of courses etc whereas student retention
means the number of students who complete their course and acquire degree
despite of any kind of circumstance and gain good grades in the transcript. In
educational institutions, rate of student attrition and retention has reached up to
considerable unit (Stallone, M. N. 2011).

Student attrition and retention not only affect the performance of education
institution or department but also affect the faculty positions and raise the
financial problems for the parents eventually. To analyze the pattern of student
retention and attrition, one has to find out the reasons of this cause then take
steps to resolve this problem. Our research topic falls into the type student
attrition and retention which is further explained in the following subsection in
detail. For this research collected dataset from Capital University of science
and Technology, Department of Computer Science. Let’s consider first
semester and second semester courses because student attrition rate is high in
first and second semester. These semesters are helpful attrition of students and
predicted the student performance.

Student performance prediction

To recognize the efforts of students in academia, there are some known
parameters carried out by universities or educational institutions (Romero, C.,
& Ventura, S, 2007). Among those parameters, student’s grades, attendance,
grades per courses are commonly evaluated. Student performance prediction
has been the growing research problem for the researchers since a decade. To
improve the institution’s prestige and efficiency with respect to student
performance improvement, there are several techniques used by different
researchers such as course specific regression, personalized linear multi
regression, classification and clustering. These approaches not only work for
educational institutions those are in existence physically but also for virtual
campuses too i-e; LMS and Moore of Stanford University (Elbadrawy, et al,

3
2016). EDM is a broad research field in which researchers are exploring a lot of
problems including student’s performance, attrition and retention, course
selection, and particular teacher’s selection. The widely covered research
problems in the domain of educational data mining are known as student’s
future grade prediction, performance prediction and course enrollment
recommender system.

This research is based on different aspects. One aspect is to identify the number
of those students who had to leave the institution because of their poor
performance in studies or financial problem. Then the number of those students
have been found who completed their degree belong to the spring semesters of
three years. Expectedly this research will help the students to improve their
results in the courses which have been taken for the experiments by
recommending them appropriate teachers and courses based on their associated
factors.

This is helpful not only for students to raise their grades but also for parents
not to suffer from financial problems. Another aspect which is being
considered is to retain maximum percentage of students in the university and
fulfill their degree requirements till its completion. Ultimately, if the grades of
students would be good their retention will become stronger in the institution.
This research might reduce the attrition rate because of the provision of teacher
recommender system, course recommender system and student performance
prediction. The outcome of our research has been discussed in detail as follows.

Students Performance Prediction is the research problem which has been

identified by many researchers with the help of different approaches such as
Matrix Factorization, classification and clustering algorithms (Polyzou, A., &
Karypis, G, 2016). Before applying these approaches, students variables are
supposed to be extracted from the collected dataset that are unique registration
number, result of previous two terms, GPA (Grade Point Average), number of
family members, family welfare, gender, usage of social networking and use of
technology and internet (Oskouei, R. J., & Askari, M, 2014). After collection,
data preprocessing is applied in which positive and negative effects of dataset

4
are removed and then classification and clustering are applied. By using such
techniques, data can be further divided into clusters and classes to analyze
useful and hidden patterns from it (Kabakchieva, D, 2013).

With respect to performance prediction of students, there are many cases such
as student’s next term grade prediction, presence performance based on
assessment in the current courses. To evaluate this problem, researchers have
used some of the characteristics of students such as admission records, High
school scores, SAT/ACT scores and grades of previously completed courses
(Elbadrawy, et al, 2016). More characteristics such as class test, seminar
attendance and marks, assignment marks (Baradwaj, B. K., & Pal, S. 2012).

Course enrollment recommender system It helps to determine the

performance of students by recommending them courses up to their
performance. For this purpose researchers have proposed the recommender `in
which he used 67 templates of active students. This recommender system not
only helps to predict the grades of students, it also helps to recommend the
courses to the students by considering their timetable (Bydžovská, H, 2013).

Proposed attributes of students affects their performance in various ways. It not

only helps in identifying the students with low academic performance but also
paves the ways to improve their grades and enables recommender systems of
universities to predict the grades of students. For this purpose, different
classification techniques of data mining have been studied in this research.
With the help of this study, student carrier can be affected by means of building
up and students can easily pave their way towards their desired destination.
This study will also help to construct the course recommender system which
will allow the students to select their courses that are recommended by the
proposed system. Another advantage of this study is highly beneficial
economically for parents. With the help of proposed system, students may not
waste their money in selection of courses to repeat. Moreover, student’s
academic record can be improved and efficiency of department management
can be raised.

5
The factors used in the research have already been applied in countries other
than Pakistan. Those countries are Canada, USA, England, and Nigeria. All
these factors have been applied in the colleges and universities of narrated
countries and now will be validated in the context of Pakistan.

1.2 Problem Statement

A major problem that the academic institutions are facing worldwide is poor
performance of students in academics. This causes high attrition rate that is a
loss to students, parents and institutions. Student’s performance prediction can
help to reduce attrition rate as it raises an early alarm for students not
performing well and are likely to leave the institution/studies. Most of the work
on performance prediction has been done in foreign institutions; this has to be
done in the context of Pakistan for more accurate results.

1.3 Research Questions

Given below is the proposed problem statement

RQ1. Whether exiting identified factors for grade prediction are valid in our
local context (Pakistan)?
RQ2. Which factors help in accurate prediction of students’ GPA of first
semester?
RQ3. Is it possible to improve the performance of a student by allocating of
appropriate teacher for a subject?

1.4 Scope
The application of this research is versatile. The experiments of this research
have been conducted on the dataset that has been collected from Capital
University of Science and Technology, Islamabad (Pakistan). Such factors can
be applied on the data of Government institutions as well as schools and
colleges.

1.5 Application of Proposed Approach

Our proposed research will be beneficial in one of the following ways in the
domain of educational data mining.

6
To make the university management system efficient
This research might be highly beneficial for the university management system
to keep check on the performance of student’s belong to different departments.
This will not only help to predict and improve student’s performance but also
supports the parents economically.

To enable Course recommender system pro-active

The selection of courses is major problem for the students. To build such
recommender system that helps students in selection and registration of the
courses will be quite beneficial for the institutions.

Helpful in improving teacher recommender system

In the proposed solution this recommender system will help department to
recommend specific teachers to particular students based on previous 0239
academic record and inclination of selected courses in which they pursued good
grades.

1.6 Significance of solution

This research is a fine contribution for the educational institutions of Pakistan.
With the help of this research, the factors that have been used by the
researchers from the foreign universities are considered in local context. These
factors are planned to apply over the data set of Capital University of Science
and Technology of Pakistan. These factors have been utilized with the help of
WEKA for the student’s performance prediction. Universities and other
educational institutions may capable of predicting the student’s next term
grades.

Moreover, with the help of the contribution of this research there are chances to
propose teacher recommender system that will help to recommend suitable
teachers to the students by considering their current performance in the
semester and their relative interest in the courses that might improve their
grades.

7
1.7 Organization of thesis
This thesis comprises of five chapters. The first chapter states the introduction
of the proposed research. Second chapter is the literature review in which the
work and research contribution in chosen area by former researchers has been
discussed. Moreover, overall literature summary has been presented in
summarized way. Third chapter contains the methodology diagram that has
been adopted to perform the experiments and how experiments will be
performed. The results of experiments by using selected tools have been
presented in the chapter four. Last chapter contains the conclusion and future
work in which summary of whole thesis has been presented.

1.8 Definitions and frequently used terms

Performance Prediction
A student’s grade is used to acknowledge his/her related performance in
academia. It is now possible to predict the student’s performance with scientific
methods in which quantitative approaches have been used by the researchers
(Bhardwaj, B. K, 2012) (Pal, S. 2012).

Recommender System
Usually recommender system is built to assign a particular entity to required
entity. In the field of data mining, researchers have proposed some
recommender systems in which teacher recommender system and course
recommender systems are involved (Bozo, J., 2010) (Alarcón, R.,2010)(
Iribarra, S., 2010).

8
Chapter 2

LITERATURE REVIEW
The initiative of online course learning system is based on e-commerce
venture. With the growth of this venture, data from web resources started
collecting and storing in the excel that contains customer and product
information and order information (Kokina, J, 2017). E-commerce is a term
that refers to use as online business through internet. There are various websites
that are working for this purpose such as e-bay, Alibaba, and Amazon etc. It
can be said that data storage in excel from the resources has been derived from
e-commerce. Predicting student’s performance by using data mining techniques
to extract information from the academic dataset of universities has become
state of the art research in the scientific society. Universities are confronting
with some challenges now a day to analyze the performance of their students.
That’s why researchers are focusing on student’s profiles and characteristics to
make the university management aware of student’s performance and overall
academic result (Kabakchieva, D, 2013). There is another dimension of
student’s performance that is the dependence of student retention upon student
student’s performance. To minimize the problem of student retention cases in
the universities, different researchers have proposed different methods to
predict the performance of students in their future semester based on the
performance of previous one.

To predict the courses of next term grades, four parameters have been
considered in this study such as; admission records, High school scores,
SAT/ACT scores and grades of previously completed courses. Based upon
these parameters, recommender system can be trained to predict the grades of
students accurately in any of the educational institution. Historical information
about the course has also been considered in this study such as which course is
taught by which teacher and information about contents of the course. Many
researchers have used LMS and Moore to predict the successive chances of
success and failure of students. In this research, regression based methods such
as course specific regression (CSPR) and personalized linear multi regression

9
(PLMR) has been used. Another method known as matrix factorization based
methods in which standard matrix factorization (MF) has been used for the
grade prediction of students (Elbadrawy, A, et al, 2015).
In this chapter, the background of educational data mining and its branches will
be discussed in detail. Also the student’s performance prediction and data
mining approaches that are commonly used by researchers in the literature are
being discussed.

2.1 Educational Data Mining

Data mining is a process of sorting data and extracting information from
existing databases.1 With the help of pattern mining and data analysis, hidden
information can be obtained from huge datasets. The strategy of data mining is
now applied in the field of education by researchers. They are busy in
exploiting a lot of dimensions in education sector. This is now known as
educational data mining. Data mining is applying in educational sector by
considering the performance of students and finding the position of students by
using their academic records. Educational dataset is being collected from
various resources such as interactive learning systems, computer-supported
collaborative systems, and administrative datasets of school, colleges and
universities2. Data mining methods are now implemented in well known
universities to analyze the patterns of student performance from the dataset
through which information can be extract and decision making may become
easier for the management of institutions (Kabakchieva, D. 2013).

With the incremental growth in the use of technology everywhere, educational

institutions are now busy in finding hidden trends and patterns in their larger
datasets (Scheuer, O. 2012). (Merceron, A. 2005). Datasets that are used for
experimental purposes in educational institutions have become possibly
available because of web based educational systems in which LMS, Moore and
Portal system have become common. With the help of these sources, dataset
can easily be collected if authorization is accessed. One purpose of extracting
information from its own dataset is to make its prestige among other

1
http://searchsqlserver.techtarget.com/definition/data-mining
2
http://www.educationaldatamining.org/

10
educational institutions stronger. Another purpose is to build the student career
by covering its each and every aspect such as improvement of their grades if
lacking, overall performance booster, support them financially, make them
enable to select the courses suggested by course recommender system, assign
them appropriate teachers based on their inclination of interest in course
selection and many more.

However, these are some common areas in which researchers are producing
their research by using different data mining techniques. In our research, same
areas are being covered under the dataset of Pakistani students. Three factors
are being applied on students those factors have been discussed briefly in
section 2.4. In current section, our focus will remain on key areas of EDM, in
which researchers are engaged.

a) Personal recommender and learning environment

Web based learning environments have become common right now. It has
become powerful medium to act as a bridge between learner and instructor and
provides interesting learning mechanisms to both (Nam, C. S, 2007) (Osmar,
R, 2002) has elaborated the concept and working of web based learning which
is also known as e-learning. A recommender system helps to recommend the
actions to a learner by using his previous actions with the help of intelligent
software agents. The use of such recommender systems were first initiated in e-
commerce and now carrying out by researchers in the domain of e-learning.
Osmar has implemented this recommender system in assisting the offline web
miners who are responsible to find hidden trends in the data and online course
navigation. With the help of recommender system, educational institutions are
now strengthening their management and departmental progress. Researchers
have proposed several recommender systems among which course
recommender system and teacher recommender system are commonly known.

b) Course management system

By means of course management, researchers have explored different sources
to find the best possible use of course management system. In this context,
Moodle is found to be open source course management system that helps

11
educators to make their courses available on the internet. All the data on the
Moodle is managed by Moodle team. Moodle is now facilitating educational
institutions, community colleges, and schools to create online teaching system
(Dougiamas, M., 2003). It delivers courses online, unlike traditional
classrooms. Data of courses and students is stored in its specified database
which is further used by researchers that carries out for mining purposes to
extract useful information from it. Performance of Moodle has raise higher in
virtual environment. Other online course management system such as web
portals and learning management systems are serving for the same purpose now
days.

With the use of internet in classrooms and institutions, those institutions have
launched their proper websites or web pages, through which students can
register themselves into particular courses (Kaminski, J. 2005). In the area of
online course learning, researchers have presented many research articles. In
this domain LMS and Moore have contributed a lot. They have been offering
different courses on their websites and thousands of students from all over the
world registered themselves. This is how, huge amount of dataset is collected
from LMS and Moore and researchers have performed analysis on these dataset
to evaluate the performance of students who enrolled (Kizilcec, R. F, 2017)
(Pérez-Sanagustín, M., 2017) (Maldonado, J. J. 2017). Self regulated learning
(SRL) is a term that has been stated by narrated researchers. According to
them, students having strong SRL are good enough in planning, managing and
controlling as compared to the students having weak SRL. In this regard,
MOOCs have been providing support for the learners with the help of different
levels of SRL.

c) Student attrition and retention

With the passage of time, growth of private educational institutions has been
increased up to the remarkable extend. These institutions have become source
of higher learning and business entity. Therefore, maximum number of
student’s enrollment is its lifeline. For the survival of private institutions,
profitability, proper management and alignment are mandatory. In this respect,
student retention until the completion of degree is quite necessary. That’s why

12
institutions are finding that factor that ultimately causes student attrition
(Azarcon Jr et al, 2014). After analyzing those factors, it is important for
educational institutions to make strategic adjustments accordingly to improve
student retention in institutions. In Gaviria, Colombia, people are closely
belongs to social mobility that is a causing the academic performance of the
students and student attrition rate may increase in educational institutions.
Researchers have used three drivers the student performance analytics. The first
driver is the volume of data that is being collected from learning management
system and student information system, second one is the e-learning and third
one is political concerns (Guarín, C. E. L., 2015)( Guzmán, E. L.,2015)(
González, F. A. 2015). The application of data mining in the field of
educational data mining has been emerged in different areas and researchers are
exploiting these areas with various dimensions.

There is another supplement for student learning through web that is known as
MOOCs. In online course learning and management system, high rate of
dropout of students have been identified by researchers. This problem has been
enlightened by the YANG in his research that is students are dropping at
considerable level in Massive Open Online Courses (MOOCs). To control the
student attrition problem from the coarser classes, researchers have proposed a
model in their research. This model is a helping hand to determine such
influential factors that are causing student drop out. So the predictors have been
tried to propose in this research that will determine the factors related to
student’s behavior and social position in the discussion in which they
participate in the forums (Yang, D.et al, 2013).

The problem of student attrition and retention is not new for the educational
institutions. It has been enlightened by the researchers from the fields of data
mining and information visualization. Now it has become very common
research problem for the researchers. Student attrition and retention problem
has been observed by the researchers when this problem was raised up to the
ratio of 50% on the colleges of Ontario (Drea, C. 2004). To reduce attrition
rates, institutions should focus on student retention. Researchers have analyzed
the factors that causes student attrition and in the research, Drea has addressed

13
both elements; student attrition and retention. To retain the persistence of
institutions, two theories have been formulated in this context. The first one is
student integration model by Tinto and student attrition model by Bean. Both
models work almost similar (Cabrera, A. F. 1993).

There are numerous reasons for what student attrition has become research
problem for the researchers. In those reasons, personal disappointments,
financial setbacks, and lowering of career and life goals are considerable.
Therefore scientists have carried out the retention and persistent in their
research to resolve this societal problem (Ramist, L.1981). This research focus
on student performance and also find a teacher methodology has positive
impact on student grade prediction and has reduced student retention rate.

2.2 Student performance prediction

Student performance prediction is the research problem which has been
considered by many researchers from the field of educational data mining
(Kabakchieva, D, 2013). With respect to performance prediction of students,
there are many cases such as student’s next term grade prediction, presence
performance based on assessment in the current courses. To evaluate this
problem, researchers have used some of the characteristics of students such as
admission records,

High school scores, SAT/ACT scores and grades of previously completed

courses (Elbadrawy, et al, 2016). More characteristics such as class test,
seminar attendance and marks, assignment marks (Baradwaj, B. K., & Pal, S.
2012). For the sake of performance prediction, data that has been stored in
open polytechnic student management system belong to 2006-2009 had been
used. For the experiments to be applied over the dataset, feature selection and
clustering techniques were applied on the dataset which are also known as data
mining techniques.

In the field of education, predicting the performance of students has become

considerably important for the university management and departments.
Because the prestige of any educational institution is ultimately depends upon
the performance of its students and if performance of maximum number of

14
students degrade up to the mark able level, it should be considered by the
department management and find out the factors that affects the performance of
students. Under consideration of this problem, researchers have evaluated their
studies against the parameters that affects the overall performance of students,
data mining techniques and data mining tools (Kaur, G, 2016)( Singh, W,
2016). In such parameters, psychological, personal and environmental factors
are involved. For the experiments, they have used Naïve Based and J48
techniques through WEKA.

a) Future grades prediction

To predict the future grades of students is the considerable research problem

that has been identified by many researchers with the help of different
approaches such as Matrix Factorization, classification and clustering
algorithms (Polyzou, A., & Karypis, G, 2016). Before applying these
approaches, students variables are supposed to be extracted from the collected
dataset that are unique registration number, result of previous two terms, GPI
(Grade Point Average), number of family members, family welfare, gender,
usage of social networking and use of technology and internet (Oskouei, R. J.,
& Askari, M, 2014). After collection, data preprocessing is applied in which
positive and negative effects of dataset are removed and then classification and
clustering are applied. By using such techniques, data can be further divided
into clusters and classes to analyze useful and hidden patterns from it. To
predict the performance of students and predict their grades, researchers have
used socio-demographic factors of the students such as age, gender, ethnicity,
education, work, status, and disability. In addition, researchers have used
environment may also affect the student’s performance that causes their
dropout or retention throughout the course accomplishment (Kovacic, Z. 2010).

b) Teacher recommender system

When students retain their degree the end of their duration, it encompasses the
quality education of any educational institution that refers to the term of student
retention. According to the national statistics in a research, fifty percent of the
students from higher educational institutions get attired. To (Bozo, J., 2010)

15
(Alarcón, R., 2010) (Iribarra, S., 2010). As online course enrollment system has
taken the place globally because of internet. In this variation in technology,
there are several web pages and websites that are busy in providing best at their
own but still contain some missing elements in their services. Therefore,
researchers have worked upon this problem and proposed solution regardingly.
To make every course learnable creatively, researchers has focused on teacher
recommendation according to the course respectively. The new recommender
system is known as A3. With the help of this recommender system, not only
accurate teacher will be associated to the relative course contents but also best
contents of the course will be available on that site after some time (Tewari, A.
S, 2015).

c) Course recommender system

A course recommender system helps to determine the performance of students

by recommending them courses up to their performance. For this purpose
researchers have proposed the recommender in which he used 67 templates of
active students. This recommender system not only helps to predict the grades
of students, it also helps to recommend the courses to the students by
considering their timetable (Bydžovská, H, 2013).

2.3 Commonly used approaches in EDM

From the literature, a lot of data mining approaches that have been carried out
by the researchers to evaluate their studies. Most commonly used approaches
have been discussed in detail in subsections of section.

a) Classification and Clustering

To improve the performance of weak students in the class, with the help of data
mining techniques, collected data of 1100 students have been transformed in
Weka. Researchers have used freeware software such as Weka, Clementine and
Rapid-Miner in this work. Different classifiers such as Naïve Based, C4.5,
Neural networks and random forest have been used. The classifiers which have
been used are Adaboost, Bagging and boosting. It is found that some factors
have same effects on both countries and some have different. Moreover, it has

16
been found that male students suffer with stress more as compared to female.
The performance of male students in Mathematics and formal SCIENCE is
better whereas performance of female students aroused better in Literature and
mnemonic SCIENCE (Oskouei, R. J.,2014)( Askari, M. 2014).In this field of
research, there exist wide varieties of benchmark which are used to evaluate the
performance and accuracy of experiments conducted by using machine learning
approaches. Different researchers have used various types of educational
datasets and each dataset is unique in its attributes (Garcia-Saiz, D., 2011)
(Zorrilla, M. E, 2011). Therefore, in the study researchers have proposed meta
algorithm to preprocess dataset. Various data mining models have been studied
in this research to find the most accurate one with the help of Meta algorithm.

Educational Data Mining approach is emerging in providing feasibility to the

educational institutions to improve their teaching and learning methodologies.
EDM is the process of applying data mining techniques in educational
institutions. In educational institutions, the data of student’s performance,
teacher’s evaluation, student’s enrollment data, and gender differences are to be
stored in the database. On this data, Classification, Clustering and association
rules are applied. These rules have been applied on mandatory courses taught in
IT department of King Saud University. To extend this approach, these rules
can be applied on elective courses as well to predict the grades of students in all
courses as well as final GPA. Furthermore, neural network and clustering can
also make these rules effective for prediction (Al-Barrak, M. A., & Al-Razgan,
M. 2016). Researchers have considered more attributes of students such as their
attendance, class test, seminar attendance and marks, assignment marks and
constructed classification tree in Weka. According to those researchers,
teachers can minimize the failure ratio of students in the semester by adopting
this approach and attributes (Baradwaj, B. K., & Pal, S. 2012).

b) Association rule mining

In the study, some factors which affects the grades of students of Iran and India
has been observed. In such factors, their respective gender, family background,
education level of their parents and their lifestyle has been encountered by

17
asking questions from them (Oskouei, R. J., & Askari, M, 2014). To evaluate
the performance of students and improve the management of educational
institutions, researchers have introduced pre university characteristics of
students which are known as student’s profile and place of secondary school,
final secondary education score, total admission score, and score achieved that
exams. In this research data of University of National and World Economy
(Bulgaria) has been collected and data mining techniques has been applied. In
such data mining techniques naïve bayes and bayes net, nearest neighbor
algorithm, and two rule learners One R and JRIP has been applied on the
dataset (Kabakchieva, D. 2013).

c) Pattern mining
As discussed earlier, it is now clear knowledge is gathered about online
teaching and learning system. Universities through internet are now big source
of student-teacher interaction and source of learning and training. This
technology has made place in the field of educational data mining in which
knowledge diffusion has become research icon for the researchers who belong
to field of data mining, web mining and graph algorithms. The data of online
teaching system is serving best to the researchers who keen to extract
knowledge from big datasets. Researchers are analyzing patterns of online
learning behavior of students and to draw outcomes from these sets of data,
they have been using machine learning algorithms and various data mining
approaches. In a study, researchers used 19,934 servers to identify the behavior
of students in Taiwan who had registered online courses and with the help of
this data, they drawn conclusion after predicting their performance too (Hung,
J. L., & Zhang, K. 2008). In addition, data mining techniques were proven to be
helpful for the course developers, online trainers, and instructional designers. In
their study, they used WEKA and KNIME tools to perform analysis of
descriptive and artificial intelligence. Moreover, for data visualization and
statistical analysis, they used

SPSS. In another study, data mining techniques such as classification,

clustering, relationship mining, prediction and social area networking have
been applied in educational data. (Sachin, R. B., & Vijay, M. S. 2012). Well all

18
of the mentioned approaches are ultimately works to represent the results after
applying over the data. The concept of data visualization has been derived from
visual reasoning (Inoue, S et al, 2017).

d) Text mining
In the past decade, there are number of tools such as WEKA, RapidMiner, R,
KEEL, and SNAPP that have been used to extract text (Useful information)
from the datasets in the field of educational data mining (Baker, R. S., &
Inventado, P. S. 2014). In other research, data mining is also known as
knowledge discovery from databases (Anand, S. S. et al, 1996). In this process,
database techniques are bind with mathematical and artificial intelligence
techniques.

2.4 Commonly used attributes for student performance prediction.

While review of papers from the literature, many data mining approaches have
been found that are applied in academic datasets of different educational
institutions for various purposes. Such approaches have been applied on the
attributes that are considered after analyzing the factors. Those factors are
briefly discussed in following passage.

I. Demographic attributes

From the literature, the importance of demographic attributes upon student’s

performance is inaugurated. According to the researchers, they condemn that
living standard of any student ultimately affect their studies and variation their
performance. Demographic factors include Age, gender, Financial status,
Balance due, Permanent address, Residential address, Guardian (Brother, uncle,
Self), Father qualification, Father Occupation, Full/ Part time student.
Demographic factors equally supports the researchers whether to find the
student attrition or student’s retention. Student’s retention is an obvious success
of educational institutions. In order to sustain from student retention and keep
the performance of students up to the mark, educational institutions better need
to recruit the limited students as per its availability. In a study, the student’s

19
general weighted average, School’s Radial Distance, and school ownership are
taken (Abaya, S. A., 2013) (Gerardo, B. D., 2013)

II. Pre-university attributes

Another factor which has been found during literature survey is pre-university
attributes. These attributes have quite great impact on student’s performance in
any educational institution. Pre-university attributes will be fruitful with respect
to accommodate the student’s interest to map in course recommendation
system. Pre-university attributes include Secondary School Grade, Higher
Secondary Grade, SAT Score, Pre-college, Pre-board, pre-program.
Researchers have used the historical data of students as well to find the attrition
rate of students from academia through cross validation process under the
classification and naïve based methods Guarín, C. E. L., 2015)( Guzmán, E.
L.,2015)( González, F. A. 2015).

III. Institutional attributes

Last factor that has been extracted from literature to improve the performance
of students is institutional attributes. These attributes will support in modeling
the teacher recommender system and future course prediction model in our
study. In institutional factors, Term 1 GPA, Term 2 GPA, Term 3 GPA, and
Term 4 GPA, CGPA, Total Credit hours taken, Total courses taken and Initial
major, current major, current enrollment status are included.

Proposed attributes of students affects their performance in various ways. It not

only helps in identifying the students with low academic performance but also
paves ways to improve their grades and enables recommender systems of
universities to predict the grades of students. With the help of this study,
student carrier can be effected by means of building up and students can easily
pave their way towards their desired destination. This study will also help to
construct the course recommender system which will enable students to select
their courses that are recommended by the proposed system. Another advantage
of this study is highly beneficial economically for parents. With the help of
proposed system, students may not waste their money in selection of courses to

20
repeat. Moreover, student’s academic record can be improved and efficiency of
department management can be raised.

2.5 Literature review summary

Table 2.1: Literature Review Summary

Used
Attributes

Reference Relevant Techniques Tools

Pre-university
Demographic

Institutional
attributes

attributes
10 fold cross
(Guarín, C. E.
Decision Trees, validation
L., Guzmán, E.
Bayesian model, Cost √
L.,González, F.
Classification sensitive
A. 2015)
model
(Abaya, S. Irecruit
Classification through
A.& Gerardo, Application √
C4.5
B. D., 2013) in UNIX
Clustering, Feature
CHAID
(Kovacic, Z. selection thorough
Model, Gian √
2010). cross validation, and
chart
CART classification
CRISP-DM, Neural
networking, Nearest
Kabakchieva,
neighbor classifier, WEKA √
D. (2013).
Rule learner, Decision
tree classifier
(Al-Barrak, M.
E-learning
A., & Al- Data Visualization,
Web Miner, √
Razgan, M. Classification
WEKA
2016).
(Baradwaj, B.
Classification and
K., & Pal, S. Manual √
Decision trees
2012)
Recommend
Matrix factorization
er system
(Elbadrawy et based methods,
based √
al, 2016) regression based
personal
methods
analytics
(Bydžovská, EDM, SNA and WEKA and
√
H. 2013). Collaborative filtering R
(Kaur, G., &
Naïve based and J48
Singh, W. WEKA √
Decision trees
2016)
(Kokina, J.,
Pachamanova, Predictive Modeling, Excel and
√
D., & Corbett, Data Visualization Tableau
A., 2017).

21
Chapter 3

RESEARCH METHODOLOGY
This chapter presents the methodology adopted to address the research
questions presented in chapter 1. The focus of the research is accurate grade
prediction of critical courses of first semester of BS CS students of CUST.
Prediction will help to take appropriate measures to control the attrition rate
and hence will be beneficial for students, parents and university. Targeting the
same objective, appropriate faculty members are also being recommended
based on the results of previous semester. The factors under consideration for
the proposed research are classified into three categories; Demographic, Pre-
university and Institutional. Such researches in the field of educational data
mining have been conducted in foreign earlier. Experiments have been
conducted in local context, Pakistan. For this purpose, all the data have been
collected from Capital University of Science and Technology, Islamabad
Pakistan. After data collection and its pre-processing, some data mining
techniques are being selected like SVM, Linear regression and Non-linear
regression in WEKA.

This study will not only be useful in improving the overall performance of
students but also reduces the attrition rate. Now it is easy to do such things on
the basis of early prediction. Field of data mining has been emerged with the
linkage of natural language processing, artificial intelligence, visual data
analytics, and social data analysis etc (Romero, C., & Ventura, S. 2013)(Zhang,
Y.et al, 2010). This early prediction is computed by using term marks and mid-
term marks of the students along with their pre-university; Intermediate/O-
levels marks and Matriculation marks and demographic factors; gender and
city. This methodology will be discussed briefly in the current chapter. The
related work of this research task has been discussed in detail in chapter 2.
Chapter 3 contains detailed discussion of proposed methodology along with the
data flow diagram that has been given as 3.1

22
 Figure 3.1: Data Flow Diagram of Proposed Methodology

Proposed data flow diagram is constructed on the basis of following set of

symbols.

Table 3.1: Description about Data Flow Diagram

This cylindrical shape represents the raw data that has

been collected for the research and need to be pre-
processed.
This filled arrow is used to represent the flow of tasks
that has been carried out for the experiments.
This symbol represents the working of pre-processing
steps in our data flow diagram.
This oval shaped rectangle represents the algorithms
and techniques that are supposed to apply on the data
set for the experiments.
This shape represents the final results that include the
result prediction and evaluation of results.

23
To conduct the experiments, first of all the data is collected from the CUST
(Capital University of Science and Technology), Islamabad Pakistan. The data
set contains the data of six semesters that belongs to the BSCS program. On the
basis adopted factors, the GPA of 1st semester as well as CGPA of 2nd semester
has been predicted. In addition, performance of teachers against their respective
courses has been measured through ANOVA analysis. The experiments have
been conducted through WEKA. Each research question has been answered
with the help of proposed methodology and experiments.

3.1 Data collection and Pre-processing

As discussed earlier, domain of computer science is being selected for the
experiments and CUST has provided the facility to collect the data from its
sources. The academia of CUST is semester based and there are two semesters
per year. One semester corresponds to the spring semester which starts from
February and ends in June. And the fall intake starts from mid of September
and ends in the end of January. The data initially contained 5 courses of first
semester, such as English-1, Physics, Calculus-1, Pakistan Studies & Islamiat
and ITC/CP.

Table 3.2: Dataset Summary

Term Registered Students

Spring 2014 74

Fall 2014 155

Spring 2015 115

Fall 2015 102

Spring 2016 62

Fall 2016 101

Spring 2017 58

Total 667

24
However, the information has been extracted from a research (Junaid. 2017),
that most crucial courses with respect to attrition or students’ performance are
ITP/CP and Cal-I. So results of only these two semesters are considered and
excluded the rest three.

Collected data is related to the students of first semesters and seven semesters
starting from Spring 2014 (Term 141) till Spring 2017 (Term 171), and also of
second semester students of six semesters starting from Fall 2014 (Term 143)
to Spring 2017 (Term 173). The courses of first semester, as mentioned earlier,
are Cal-I and ITC/CP, and the second semester courses are Cal-II and CP/OOP
from each spring new intakes that belongs to the semester no 1. The initiative
of this research is to identify the factors which contribute to keep the students
retained till the end of the semesters as well those factors that affect the
performance of the students. Following table 3.2 shows the information of the
students comprehensively.

Table 3.2: Summary of Session-wise Attrition

2nd Sem

3rd Sem

4th Sem

6th Sem

7th Sem

8th Sem
1st Sem

5h Sem
Term

141 64% 20% 8% 8% 0% 0% 0% 0%

143 29% 35% 25% 12% 0% 0% 0% 0%

151 60% 30% 9% 0% 0% 0% 0% 0%

153 71% 25% 4% 0% 0% 0% 0% 0%

161 94% 6% 0% 0% 0% 0% 0% 0%

163 100% 0% 0% 0% 0% 0% 0% 0%

Attrition % 70% 19% 8% 3% 0% 0% 0% 0%

25
The dataset contains the institutional factors; Registration Number, Name,
GPA, Internal marks like midterm, and term works marks etc and overall
information of the students such as Student’s registration number, GPA,
Registered courses, Matriculation and Intermediate results. To conduct the
experiments, we have used demographic and pre-university factors of the
students apart from institutional factors. In the data set, the data of term 141,
143, 151, 153, 161, 163, and term 171 are included. The terms of the spring
semesters are associated with the registration numbers of the students such as
the term 141 means spring 2014. Similarly, the data of spring 2015 and spring
2016 are collected and on the basis of three years record of the students. The
file of dataset is initially imported into WEKA after converted into CSV and
ARFF format for the pre-processing and further experimentations have been
conducted respectively.

Missing Data Handling

The problem of missing data generally arises due to absence of data in an

observation for any variable during experiments. Missing data problem also
arises when no information is provided or unavailable for the variables. The
reason of missing data may involve the problem in collection of data by the
researchers or the sources have not provided it completely3.
Missing data has been handled by two ways. Firstly, null values of courses
have been found from other semesters and may be read in late semester in
which students have passed that course. In case of consistent absence of values,
the grades of such course have been supposed by taking the average result of
other courses in respective semester. The missing data can be handled through
different filters in WEKA.

3
https://measuringu.com/handle-missing-data/

26
Noisy Data Handling
While data collection, irrelevant data is called as noise. As the data of BSCS
students of spring 2014 from spring 2016 is collected, therefore during the
collection, data belongs to Software engineering and bioinformatics was
present and then eliminated through filters in WEKA. Noisy data occurred in
the form of errors such as; GPA = “−0.4” and Intermediate or Matriculation
marks = “-455, 544.8” which were handled in Pre-processing of data.

Attribute Selection
After missing data handling and outlier detection phase, final attributes have
been selected on which overall experiments and results depend. Such attributes
have been narrated in table 3.1 that belongs to all three types of factors;
demographic, pre-university, and institutional. One more factor has been
considered to answer our research question that is teacher performance. After
completion of data pre-processing, classification algorithms have been applied.
Attributes are then finalized for the experiments that have been evaluated to
answer the research questions. In the current section, we have discussed
adopted classifiers in detail.

3.2 Classification
Before applying algorithms, factors have been grouped that are earlier used by
foreign researchers. These factors have been compared against the factors used
by the researchers in local context (Pakistan). Moreover, we have expanded our
research work by making the variations of different combinations attributes and
then compared. The comprehensive detail about these factors has been
mentioned in table 3.4.

27
 Table 3.3: Selection of Attributes
Sr.No. Factors Attribute Name
1. Gender
2. Age
3. Residence
4. DEMOGRAPHIC Location
5. Race
6. Father’s qualification
7. Father’s occupation
8 Secondary school grade
9 Higher secondary grade
10 PRE-COLLEGE Pre-college
11 Pre-program
12 SAT score
13 GPA (Term1,
Term2,Term3, Term4)
14 CGPA
15 Financial status
16 Total credit hours taken
INSTITUTIONAL
17 Total courses taken
18 Initial Major
19 Current Major
20 Current enrollment status
21 Teacher methodology
22 Instructor grading
23 Instructor feedback
24 TEAHCER Instructor teaching
methodology
25 Teacher name

28
3.3 Algorithms and Techniques
To evaluate our research by using different classification algorithms; Naïve
Based and J48 has been compared. These classifiers are chosen on the basis of
different reasons i-e; these classifiers supports Categorical and Nominal. For
this purpose WEKA 3.8 tool has been used. The tool is open source available
on the web and is generally used for machine learning algorithms,
classification, training and testing of data. The data is then can be visualize in
the form of graphs as well. In present research, the formed combinations of
attributes have been evaluated to answer our research questions. For this
purpose the data has been converted into csv format to import in the WEKA
and the results from various filters and classifiers have been accumulated.

a) Naïve Based
Naïve based algorithm is comparatively fast algorithm in terms of
classification. It works faster on huge datasets by using Bayes algorithm of
probability. Bayes algorithm generally used to predict the class of unknown
dataset4. Naïve based algorithm works on assumptions to label an item whose
features are known but name is unknown. For example; a fruit is labeled as an
apple if it is round and red in color and its size is 3 inches in diameter. These
features of apple will raise the probability of this fruit that it is an apple.

b) J48
J48 decision tree is used to predict the target variable of new dataset. If dataset
contains predictors or independent variables and set of target or dependent
variables, then this algorithm is applied to extract the target variable of new
dataset5.

c) Linear Regression

Generally linear regression classification algorithm is an approach to identify

the relationship between dependent and independent variable. It is generally
used for predictive analysis and has two main points. One is to check whether
predictor variable does a good job in predicting the expected outcome variable.

4
https://www.analyticsvidhya.com/blog/2015/09/naive-bayes-explained/
5
http://data-mining.business-intelligence.uoc.edu/home/j48-decision-tree

29
Second main thing that linear regression does is the identification of variable
that are significant predictors of dependent variables. At the end, the regression
equation is used which helps to determine the set of predictor which are used to
predict the outcome. In this research, algorithms are being used to compare the
trend and pattern of the factors with other approaches like non-linear regression
and SMO.

3.4 Evaluation of Research Questions

For our research, we have proposed three research questions that have been
evaluated in multiple dimensions. Such dimensions have been explored in
following ways respectively.

RQ1: Whether exiting identified factors for grade prediction are valid in
our local context (Pakistan)?
Our research question number 1 determines the validity of factors in local
context which have been carried out in foreign context by the researchers. The
main reason that distinguishes this research in local context from the foreign
context is that the norms, traditions and culture of every country that varies.
Therefore it is needed to be evaluated that whether the existing factors that
work in the colleges and universities of Canada, USA, England, and Nigeria
etc, are applicable in Pakistan or not. After pre-processing of whole data,
classification algorithms have been applied to analyze the behavior of
demographic, pre-university and institutional factors. According to the results,
it has come to known that these factors are also valid in Pakistan because of
nearest difference in accuracies. Detailed results have been discussed in chapter
no 4.

For our research, data being collected from term 141 to term 171 which consist
of 667 students. Then three types of exiting attributes to predict the
performance of students have been taken. First type of attribute is
demographic; Gender, Age, Residence, Location, Race, Father’s Qualification
and have considered only Gender and Age of the student. Second type of
attribute is Institutional; GPA(Term1, Term2, Term3, Term4, CGPA, Financial
status, Total credit hours taken, Total courses taken, Initial Major, Current

30
Major, Current enrollment status, Teacher methodology and Midterm was
considered, Term Marks for the prediction of final CGPA.

Third type of attribute is Pre-qualification; Secondary school grade, higher

secondary grade, Pre-college, Pre-program, SAT score and Marks of FSC and
MATRIC are considered. Among these attributes, specific attributes

During pre-processing, it has been found the total number of students who
could not continue their program and left the course incomplete. After pre-
processing, there were total 660 remained and 7 out of 667 who left the course.
Two types of errors occurred in this dataset. The first was missing data and
second was incomplete data which has been corrected in the phase of pre-
processing.

In Table 3.5, Attributes selection to dataset for experiments has been shown

Table 3.5: Attributes selection to Dataset

Mid Physics
Mid Eng-1

Mid Cal-1
Mid ITC
Gender

Studies
Sr. No.

Matric

GPA
F.Sc
City

1 Male Rawalpindi 51 64 10 10 16 9 16 2.51

2 Male Islamabad 61 56 9 9 8 8 8 2.22

3 Male Rawalpindi 51 64 7 7 13 17 13 3.15

4 Male Islamabad 48 55 9 9 12 7 12 3.26

5 Male Islamabad 77 58 12 12 17 18 17 3.07

6 Male Islamabad 74 60 9 9 16 18 16 2.74

7 Male Rawalpindi 84 70 6 6 12 7 12 2.86

8 Male Islamabad 64 51 7 7 12 6 12 1.9

9 Female Islamabad 61 56 8 8 11 6 11 1.83

31
 RQ2: Which factors help in accurate prediction of student’s GPA of first
semester?
Some specific attributes form all three factors; demographic, institutional and
pre-university have been selected. The combinations of those attributes are
constructed and then filters have been applied respectively. There are total 21
unique combinations of attributes presented in table 3.6.

Table 3.6: Pattern of Combination of Attributes

Sr.
No. ATTRIBUTES
Demographic, Pre-Qualification, Institutional (Result all subject of 1st
1 semester)
Demographic, Pre-Qualification, Institutional (Result Eng-1, Cal-1, ITC,
2 subject of 1st semester)
Demographic, Pre-Qualification, Institutional ( Result Eng-1, Cal-1 subject
3 of 1st semester)
Demographic, Pre-Qualification, Institutional (Result Eng-1 subject of 1st
4 semester
5 Demographic, Pre-Qualification, Institutional (Cal-1 subject of 1st semester)
Demographic, Pre-Qualification, Institutional (Result ITC, subject of 1st
6 semester)
Demographic, Pre-Qualification, Institutional (Result Cal-1, ITC subject of
7 1st semester)
Demographic, Pre-Qualification, Institutional (Result Eng-1, ITC subject of
8 1st semester)
9 Demographic, Pre-Qualification, Institutional(GPA )
10 Demographic, Pre-Qualification
11 Demographic
12 Pre-Qualification
13 Pre-Qualification, Institutional (Result all subject of 1st semester)
14 Demographic, Institutional ( Result all subject of 1st semester)
15 Institutional (Result Eng-1, Cal-1, ITC, subject of 1st semester)
16 Institutional (Result Eng-1 subject of 1st semester)
17 Institutional ( Result Cal-1 subject of 1st semester)
18 Institutional (Result ITC, subject of 1st semester)
19 Institutional (Result Eng-1, Cal-1 subject of 1st semester)
20 Institutional (Result Cal-1, ITC subject of 1st semester
21 Institutional (Result Eng-1, ITC subject of 1st semester)

32
RQ3. Is it possible to improve the performance of a student by
allocating of appropriate teacher for a subject?
There is a fact that is normally faced during the research period is; the
personality of teacher affects the performance of the students. If the
background of the teacher is already known with the help of prediction system,
it might become feasible for the department to recommend the best suitable
teacher to the student up to his level of interest. And this phenomenon can
surely boost the performance of the student. To answer the question number 3,
experiments have been performed using two approaches have been used. And
these approaches Statistical experiments and Predictive experiments.

a) The research has used combination of techniques comprising simple

mean and the prediction based on teacher. The focus of this question is
to find the most suitable teachers for Cal-I and ITP as these two subjects
are considered the first semester with respect to the academic
performance or attrition rate. Following calculations have been
performed for the Cal-I teachers with respect to the prediction
experiments:

1. Average of prequalification is taken

2. Average of actual Cal-I marks based on the Cal-I teacher

3. Average of actual Cal-II marks based on the Cal-I teacher

4. Prediction of the Cal-I and ITC/ITP is taken on the basis of

prequalification

5. Predict Cal-ll marks on the basis Cal-I and average is taken

b) ANOVA data analysis

ANOVA is a statistical technique which is used to measure the difference in a
scale level dependent variable by a nominal variable which comprises of two or
more types of categories. Normally ANOVA test is applied to find out the
significant difference among groups. Similarly we will be applying this test to
find the difference between the performances of teacher’s subject wise in our

33
case of experiments6.
c) Z-score data analysis
Unlike ANOVA (analysis of variance), Z-score is usually applied on three or
more means. A Z-score is a type of hypothesis test which is a way to find
whether the results obtained from a test are valid or need to be repeated. For
example, if someone said they had found a new drug to cure the cancer, one
would want to be sure it was probably true. Similarly, in our research, we have
applied Z-test to compare the performance of teachers particularly with respect
to their subject. Z – Test will exploit the likelihood that the obtained results are
true or not. A Z-test is generally used when the data is approximately normally
distributed in the form of pairs7.

6
http://www.statisticssolutions.com/manova-analysis-anova/
7
http://www.statisticshowto.com/z-test/

34
Chapter 4

RESULTS & EVALUATIONS

Researchers have made their significant contributions to introduce such factors

in the field of educational data mining that are quite beneficial for the
educational institutions to mine hidden patterns from the academic database.
Educational data mining is broadened field in which Online learning
management system as well as physical institution’s datasets are used. One
limitation of this approach is to get the access of the dataset from the
administration in order to perform analysis and conduct research. The dataset
have been collected from Capital University of Science and Technology,
Islamabad and applied existing factors on this dataset to answer our proposed
research questions. Following sections of this chapter will present the overall
picture of obtained outcomes in detail.

Data collection and pre-processing are the initial steps towards the analysis of
the research. For this purpose, the demographic data has been collected from
the registrar office of the CUST which contains student’s gender and city. With
the co-operation of administrative resources of CUST, the data related to pre-
university and institutional factors has been collected. Institutional and pre-
university factors contain the data about mid-term marks, term-marks, GPA,
intermediate marks and matriculation mark. University portal provided
privilege to access this data of the students to gather the sufficient dataset for
our experiments.

The gathered data came into many formats such as PDF, word and excel. For
the pre-processing, firstly this data set brought into the single file; Excel.
Figure 4.1 represents the format of files and transformation of all files into
single one. The collected data set arises in three formats; word, PDF and excel.
After collection of dataset, it became mandatory to transfer the data into single
file from all the files.The purpose of excel file is to import it into the WEKA to
apply filters for the pre-processing. The final versions of dataset in the form of

35
excel is then converted into CSV format and imported in to the WEKA. Such
filters have been applied that are discussed in subsections of section 4.1.

Figure 4.1: Formats of Data

In context of missing data, lacking attribute values, lacking certain attributes of
interest, or containing only aggregate data instead of complete information are
found. Incomplete information may lead the research towards in accurate
outcomes that might affect the overall experiments. Therefore, the missing data
problem has been handled in a way that no attribute remained null or
incomplete.

Table 4.1: Occurrence of Missing Values

Final Exam

Mid Term
Weighted

Teacher
Gender
Sr. No.

Work
Term
Total
ITC

1 Male B- 75.95 76 25.2 15 21.67 D

2 Male D 50.27 50 17.6 7.25 14.8 C

3 Male C 66.93 67 22.8 8.25 21.65 A

4 Male

5 Male B- 76.43 76 21.6 13.75 27.27 D

6 Male D 50.19 50 14.4 10.25 10.95 B

7 Male A 89.93 90 34.8 14.25 30.57 D

36
The methods of data mining behave differently in the way that they treat
missing values. Normally, they ignore the missing values, or exempt those
records which contain missing values or either replace missing values with the
mean, or conclude missing values from existing values. Missing
Values Replacement Policies include the number of strategies such as; ignore
the records with missing values. Following example shows the occurrence of
missing values in the dataset. After the handling of missing data by taking the
average of last three records of students and filled the missing one, complete
dataset is gained. Further, this dataset has been used for experiments to
evaluate all research questions.

Table 4.2: Handling of Missing Values

Term Work
Final Exam

Mid Term
Weighted

Teacher
Gender
Sr. No.

Total
ITC

1 Male B- 75.95 76 25.2 15 21.67 D

2 Male D 50.27 50 17.6 7.25 14.8 C

3 Male C 66.93 67 22.8 8.25 21.65 A

4 Male C+ 72.91 73 19.2 18 27.75 A

5 Male B- 76.43 76 21.6 13.75 27.27 D

6 Male D 50.19 50 14.4 10.25 10.95 B

7 Male A 89.93 90 34.8 14.25 30.57 D

Outliers are an observation point that is found to be distant from other

observation points. An outlier may occur due to the change in the measurement
or it may indicate experimental error; the latter are sometimes excluded from
the data set. After handling noise from the dataset, following result is obtained.

37
Attribute Selection

Appropriate selection of attributes is very important task in the process of pre-

processing. It ultimately affects the performance of results and may lead to
inaccurate results, if not chosen carefully. That’s why we have classified the
attributes into two types; Nominal and Categorical. On these two types of
attributes researcher have selected the attributes that best suit to the nature of
these two types of attributes.

Sampling
Data sampling is a statistical analysis technique used to select, manipulate and
analyze a representative subset of data points in order to identify patterns and
trends in the larger data set being examined. For example: term
141,143,151,153,161,163 dataset has been used as a training data and term 171
is test data because prediction of result of corresponding students is required to
be computed.

4.1 Final Attribute Selection

From the literature, there are plenty of attributes found and some of them have
been stated in the following table. This table belongs to the table presented in
chapter no 3 but used in the current chapter with the addition of column named
as selected attributes. These attributes have been marked to distinguish from
each other. Then final attributes have been acquired for the experiments to
fulfill the requirements of our results.

4.2 Result of Evaluation of Research Questions

With the help of adopted algorithms for the experiments in WEKA, the
proposed research questions have been tried to answer. These answers make
this research methodology validate and complete. Classification and linear
regression has been selected to measure the analysis. The results of
experiments have been discussed in detail in this section for each research
question.

38
RQ1: Whether existing identified factors for grade prediction are valid in
our local context (Pakistan)?
As existing factors along with their attributes have been stated in above
sections, we have drawn the combinations of local factors with the existing
factors. We found a slight variation between both types

of attributes. The comprehensive results of such variation have been arranged

as the answer to each of three questions. The format of attributes has been
presented in table 4.3 which are used in experiments.

Table 4.3: Attributes table

Mid Pak-Studies
Mid Physics
Mid Eng-1

Mid Cal-1
Mid ITC
Gender
Sr. No.

Matric

GPA
City

FSc

1 Male Rawalpindi 51 64 10 10 16 9 16 2.51

2 Male Islamabad 61 56 9 9 8 8 8 2.22

3 Male Rawalpindi 51 64 7 7 13 17 13 3.15

4 Male Islamabad 48 55 9 9 12 7 12 3.26

5 Male Islamabad 77 58 12 12 17 18 17 3.07

6 Male Islamabad 74 60 9 9 16 18 16 2.74

7 Male Rawalpindi 84 70 6 6 12 7 12 2.86

8 Male Islamabad 64 51 7 7 12 6 12 1.9

9 Female Islamabad 61 56 8 8 11 6 11 1.83

As an input we have imported the .CSV into the WEKA that contains
information about each attribute against every student. The input data contains
the data of students of BSCS department from term 141 to term 163. The

39
dataset of CUST (Capital University of Science and Technology) has been used
for this purpose. Among those factors, attributes like Pervious Course
Marks/Grade, GPA, SSG, HSSG, Gender, City has been used. Afterwards, we
have applied well-renown classification algorithms i-e; Naïve Bayes and J48 on
the dataset.

Table 4.4: Classification of attributes and algorithms

Reference Attributes Class Label Naïve Bayes J48

Student
Demographic, Per-
Qualification, 1st
semester 5 Subjects
Grade CGPA Grade 69.12% 72.39%

Student
Data mining Demographic, High
approach for School Background,
predicting student Scholarship, Social
performance (2012) network interaction CGPA Grade 76% 73%

Comparison of Midterm marks all

Classification subjects GPA Grade 65.6% 80.98%
Techniques for
predicting the
performance of
Students Academic Internal assessment,
Extra-Curricular
Environment(2014) activities GPA Grade 66% 73%

Student
Predicting Student Demographic, Per-
Performance: A Qualification, CGPA Grade 49.90% 63.19%
Statistical
Student
and Data Mining Demographic, High
Approach(2013) School Background CGPA Grade 50% 65%

In table 4.4, we have summarized the attributes of each factor with

corresponding to their accuracies. These attributes have been assigned with
class label and the reference of each attribute has been given as well. During
the literature review, we have found the accuracy of each attribute in their
respective scientific research publication and compare the accuracy of our

40
results. The inferences show the slight variations which is quite considerable
with the correspondence to our research. The existing attributes have been
shown in black color in the summary table and attributes in red color are local
attributes.

The summarized results have been shown in the figure 4.2 in which bars of
different colors on X-axis shows the presence of attributes and y-axis shows the
interval of accuracy.

Figure 4.2 Comparison of local and Existing attributes

Our Findings narrate that existing identified factors for grade prediction are
considered to be comparable with the local factors because of low variation in
the result of their accuracies. According to the results, Pervious Grade, Internal
Marks, HSSC were found to be effective for grade prediction.

RQ2. Which factors help in accurate prediction of student’s GPA of first

semester?

Our second research question is the extraction of those factors that helps in
accurate prediction of GPA of students belong to 1st semester. To answer the
research question, we have transformed this question into following two ways.

41
 1. Based on Mid term marks

In this phase of experiment, we have combined the attributes including

marks of mid term to predict the GPA of the first semester. our prediction
leads to the positive focal point where our proposed combinations worked
well.

2. Based on Pervious Grade

In this phase of experiment, we have interpreted our attributes to predict the

CGPA of the students. The CGPA of 2nd semester has been computed on the
basis of previous academic record.

After compilation of attributes, their accuracies have been ranked that is

presented in table 4.5 is descending order.

Information Gain Ranking Filter

Table 4.5: Ranked Attributes

Ranked Attributes
0.2277 Cal-I MID
0.1903 Cal-II MID
0.1778 ITC MID
0.1684 Physics MID
0.1548 Pak-Studies MID
0.1514 FSC
0.1006 City
0.0992 Matric
0.0791 Gender
According to results, picture states the prediction of CAL-I MID and CAL-II
MID remained better in terms of accuracy. Then the accuracies gradually
decreases from ITC MID to GENDER. Based on the corresponding accuracies
of attributes, we may inference the importance of acquired ranking information
gain of all the instances.

42
1- First Semester GPA Prediction

We have expanded our answer for the research question number 2. With the
help of “GPA Prediction for the first semester” we became able to evaluate our
results through WEKA. We covered the list of attributes narrated in table 4.6 in
which the combination of attributes are defined in column 1. There are four
combinations of attributes which belong to the three factors; Pre-Qualification,
Demographic and Institutional. The performance of J48 is considerably better
then Naïve Bayes that has been shown in table as well as presented in figure
4.3.

Table 4.6: GPA Prediction for the first semester

Attributes Class Label Naïve Bayes J48

Pre-Qualification, Institutional (Internal

marks only Midterm) all subject of 1st
semester GPA Grade 64.62% 81.39%

Institutional (Internal marks only

Midterm all subject of 1st semester) GPA Grade 65.64% 80.98%

Demographic, Pre-Qualification,
Institutional (Internal marks only
Midterm all subject of 1st semester) GPA Grade 66.46% 79.55%

Demographic, Pre-Qualification,
Institutional (Internal marks only
Midterm Eng-1, Cal-1 ,ITC subject of
1st semester) GPA Grade 69.33% 78.12%

According to the results, the performance of Demographic, Pre-Qualification,

Institutional (Internal marks only Midterm Eng-1, Cal-1, ITC subject of 1st
semester) remained better as compared to the other three combinations of
factors when Naïve Bayes classifier applied. On applying J48 classifier, the
performance of Pre-Qualification, Institutional (Internal marks only Midterm)
all subjects of 1st semester remained better comparatively to predict the grades
of first semester.

43
 Figure 4.3: GPA Prediction for the 1st semester
2- CGPA of second Semester

To predict the CGPA of 2nd semester, we have defined the following narrated
set of experiments as summarized in table 4.7. We have merged all the grades
of previous semester along with the adopted factors. These combinations and
their accuracies on the basis of Naïve Bayes and J48 have been presented in
table 4.4.

Table 4.7: CGPA Prediction for the second semester

Class Naïve
Attributes Label Bayes J48

Demographic, Pre-Qualification,
Institutional(Subject Grade Cal-1 subject of CGPA
1st semester) Grade 67.08% 74.44%

Demographic, Pre-Qualification,
Institutional(Subject Grade ITC subject of 1st CGPA
semester) Grade 68.10% 74.44%

Demographic, Pre-Qualification, Institutional

(Subject Grade Cal-1, ITC subject of 1st CGPA
semester) Grade 70.55% 73.82%

Demographic, Pre-Qualification, Institutional

(Subject Grade Eng-1, Cal-1 subject of 1st CGPA
semester) Grade 65.44% 72.60%

44
These tabular results have been presented as figure 4.4. Like 1st semester GPA
prediction, we have used same classifiers for the CGPA prediction for 2md
semester. The only change lies in the combinations of attributes. According to
the accuracy of each combination, we reached to the point that, the accuracy of
Demographic, Pre-Qualification, Institutional (Subject Grade Cal-1, ITC
subject of 1st semester) was high whereas when J48 classifier was applied, the
prediction of Demographic, Pre-Qualification, Institutional(Subject Grade Cal-
1 subject of 1st semester) and Demographic, Pre-Qualification,
Institutional(Subject Grade ITC subject of 1st semester) remained better than
other two combinations with the minor difference i-e; Demographic, Pre-
Qualification, Institutional (Subject Grade Cal-1, ITC subject of 1st semester)
and Demographic, Pre-Qualification, Institutional (Subject Grade Eng-1, Cal-1
subject of 1st semester).

Figure 4.4: CGPA Prediction for the second semester

Our Findings enlighten the following assumptions.

 For Grade Predictions, the Attribute Evaluator has computed the

Information Gain and ranked the accuracy of each attribute accordingly.

45
 Then it is found that the subjects of First semester i-e; Cal-1, Eng-1,
ITC have high value of information gain and played an effective role in
order to predict the accurate grades of students.

 As we came to know from the situation of dataset that the Attrition rate
is high in first semester. The purpose of the research is to reduce the
attrition rate, so we have considered these subjects to predict the
attrition rate of students as well as student’s performance.

 Results states that Demographic attributes does not play significant role
in grade prediction of students.

 In addition, internal marks like Mid Terms are used for early grade
prediction that has high accuracy and also helps to minimize the
attrition rate.

RQ3. Is it possible to improve the performance of a student by allocating of

appropriate teacher for a subject?
The RQ3 has been addressed with a combination of techniques comprising
simple mean and the prediction based on teacher. As it has been discussed in
the precious chapter that the focus of this question is to find the most suitable
teachers for Cal-I and ITP as these two subjects are considered to be critical for
the first semester with respect to the academic performance or attrition rate. To
evaluate the research question no 3, we have adopted the following set of
experiments. Initially, the dataset has been imported into WEKA that contains
the information of students from term 141 to term 163. The set of attributes
consist of the Pervious Core and Elective Course Internal Marks, GPA, SSG,
HSSG, Gender, City. After finalizing the set of attributes, we have applied
following proposed techniques.

• Predictive Approach: Classification algorithm: Linear Regression

• Statistical Approach: Simple average , Z-Score analysis , ANOVA

data analysis

The experiments have been conducted subject wise for the individual course
grade prediction. On the basis of course grades, accumulated teacher

46
performance has been computed. Furthermore, following calculations have
been performed for the Cal-I teachers:

1. Average of prequalification is taken

2. Average of actual Cal-I marks based on the Cal-I teacher

3. Average of actual Cal-II marks based on the Cal-I teacher

4. Prediction of the Cal-I is taken on the basis of prequalification

5. Predict Cal-ll marks on the basis Cal-I and average is taken

Pre-Qualification vs Maths Courses

The comparison of Maths course and Pre-qualification has been made in this
phase of experiment. The teachers have been labeled alphabetically and the
overall average of the students has been given against every teacher in table
4.8.

Table 4.8: Pre-Qualification VS Maths Courses

Teacher Pre-Qualification average Cal-I average Cal-II average

A 58.37142857 58.7429 64.4269

B 61.70231214 64.8324 68.03595

C 59.92207792 57.91 66.5098

D 60.17073171 56.1951 64.18938

E 58.96428571 73 72.75

F 58.95588235 57.2059 54.42167

G 60.73512748 61.8491 63.43205

H 59.82369942 61.3628 68.46664

The result shows that the performance of teacher “B, D and E” is found to be
better in Pre-qualification combination. With respect to the Cal –I performance
of teacher “B, G and H remained better. On using the Cal-II attribute, the
performance of teacher “B, C, and D” found to be better than other teachers.
These results have been shown graphically in figure 4.5.

47
 Figure 4.5: Pre-Qualification vs Maths Courses
Average of Cal-ll Teachers
In this case, we have predicted the the performance of teahcers who teach Cal-
II and compared with the actual performance of the teachers in Cal-II.
According to the results, we have found significantly unexpected outcomes.
There is a minor variation between actual Cal-II and predicted Cal-II. The
results have been narrated in table 4.9 and presented in figure 4.6.

Table 4.9: Average of Cal-ll teachers

Teacher Cal-II ave Pred-Cal-II ave

A 64.42689655 62.16017221
B 68.03594595 69.39823039
C 66.50980392 63.70893212
D 64.189375 64.42028569
E 72.75 71.255395
F 54.42166667 56.0806705
G 63.43205128 67.83774562
H 68.46663551 67.88178679

48
 Figure 4.6: Average of Cal-ll teachers
Average of Cal-I teachers

The average of Cal-I teachers has been compared with the results of actual and
predicted values. There is another attribute “Pre-qualification” that has been
used in this case. The obtained results have been given in the table 4.10 as well
as presented in figure 4.7.

Table 4.10: Average of Cal-I teachers

Teacher Pre-Q comb Pred-C1-Preq Cal-I ave

A 58.37142857 60.6696821 58.7429
B 61.70231214 61.50918169 64.8324
C 59.92207792 61.91539229 57.91
D 60.17073171 59.51882925 56.1951
E 58.96428571 60.379367 73
F 58.95588235 59.785871 57.2059
G 60.73512748 63.12452977 61.8491
H 59.82369942 61.70135392 61.3628

49
 Figure 4.7: Average of Cal-I teachers

Average of Maths Teachers

Like cal-I and Cal-II the overall performance of math’s teachers have been
computed against each teacher who teaches Cal-I and Cal-II. The information
in table 4.11 narrates that the performance of teacher “B and E” remained better
than other teachers who teach same subjects. This information has been
presented in figure 4.8.
Table 4.11: Overall Average of Maths Teachers

Teacher Over All Performance

A 60.87421589
B 65.09561403
C 61.99324125
D 60.89886433
E 67.26980954
F 57.2899981
G 63.39571083
H 63.84725513

50
 Figure 4.8: Overall Average of Maths Teachers

After the computation of results of Cal-1 and Cal-II, we became able to find the
teacher “E” whose performance found to be better and able to recommend
teaching Cal-I and Cal-II. Similarly, we will compute the results of other
subjects which are to be taught in the 1st semester and then computed similarly.
Firstly, researchers have computed the performance of every teacher of each
subject which is being taught in first and second semester and gained the
results. The results have been interpreted for every subject with respect to its
related teacher. These results have been shown significant average of all
students in Cal-1 & Cal-2 with Cal-1 teacher name. After the computation of
results for Cal-1 and Cal-II, it became clear to find that the teacher “E” whose
performance found to be better and able to recommend teaching Cal-I and Cal-
II then teacher B have better performances and further teacher B and H
performs better among all other teachers.

Further the average of these teachers will be checked whether it is significant or

not significant by applying statistical test “ANOVA”.

ANOVA data analysis on Cal-I

ANOVA analysis is used for group data we have many teacher for Cal-1 &
Cal-2 subjects. The prediction shows that teacher average performance is

51
significant. To prove these results, we applied ANOVA analysis on this dataset.
We have applied ANOVA (Analysis of Variance) test on teacher’s average. By this
way, results indicate the performance of all Cal-I teacher on the basis of group.

Table 4.12: ANOVA Analysis on Cal-1

Anova: Single Factor

SUMMARY

Groups Count Sum Average Variance

Row 1 5 303.5967 60.71935 15.95621

Row 2 5 330.7894 66.15789 3.27359

Row 3 5 313.1846 62.63691 19.02353

Row 4 5 292.9438 58.58876 27.08998

Row 5 5 357.9515 71.5903 2.338546

Row 6 5 274.268 54.85359 20.50589

Row 7 5 313.5665 62.7133 1.073095

Row 8 5 318.981 63.79621 18.29478

ANOVA

Source of
SS Df MS F P-value F crit
Variation

Between
872.7731 7 124.6819 9.273853 3.14E-06 2.312741
Groups

Within Groups 430.2225 32 13.44445

Total 1302.996 39

In this case F > F crit has been shown as 9.273853 >2.312741. Researchers
reject the null hypothesis and concluded that there are significant differences
between the methods because all 11 methods don’t have the same mean.

52
Z-Score Data Analysis
Now Z-score test is applied to evaluate the performance of every pair of
combination of teacher’s subject wise. Generally, Z-test is applied on two
populations to compare its proportion. In this case, we are supposed to compare
the performance of teachers of the same course which is to recommend to the
student. Researchers have classified each Z-test with respect to the subject and
then have considered the results of Z-score with respect to Cal-1. All the
teachers who teach Cal-I have been indicated with alphabetic characters i-e; A,
B, C, D, E, and F. Pairs such as teacher A with other teachers of Cal-1 are
made for experiment. Moreover, for the comparison, teacher’s data has been
classified as same number of students and different number of students.

Table 4.13: Two Samples for Mean with respect to teacher “E”

z-Test: Two Sample for Means

Variable Variable
1 2
Mean 74.94286 58.74286
Known Variance 102.1429 333.391
Observations 15 35
Hypothesized Mean
Difference 0
Z 4.008258
P(Z<=z) one-tail 3.06E-05
z Critical one-tail 1.644854
P(Z<=z) two-tail 6.12E-05
z Critical two-tail 1.959964

Z-score data Analysis applied on the Cal-1 teacher “E” with Teacher “B” on
the basis different number of student. Value of Z-critical to tail < Z concluded
that the performance of teacher “E” is not significant.
Then teacher “B” has been paired with teacher “A” on the basis same number
of students. The value of Z-critical > Z concluded that significant performance

53
of teacher “A”. The results have been shown in table 19. Z-test has been
applied in both cases and results have been compared either the performance of
teachers remain same or vary. By this way, results indicate the performance of
every teacher of Cal-I on the basis of combinations. The result of z-test will
evolve around this statement “If z< z critical two tail then performance
difference is not significant”. If z > z critical two tail then performance
difference is significant.

Table 4.14: Two samples for means with respect to teacher “B”

z-Test: Two Sample for Means

Variable Variable
1 2
Mean 74.94286 58.74286
Known Variance 102.1429 333.391
Observations 15 35
Hypothesized Mean
Difference 0
Z 4.008258
P(Z<=z) one-tail 3.06E-05
z Critical one-tail 1.644854
P(Z<=z) two-tail 6.12E-05
z Critical two-tail 1.959964

Here, researchers have applied Z-score data Analysis on the Cal-1 teacher “B”
with Teacher “A” on the basis different number of students who enrolled in
different terms from term 141 to term 171. In this case, value of Z-critical two
tails is greater than the value of Z which shows that performance of teacher “B”
is not significant.

Then teacher “B” has been paired with teacher “A” on the basis same number
of students and Z-test has been applied. The results shows that value of z-
critical shows the significant performance of teacher “A” as it is greater than
the value of Z.

54
Teacher B with different and same number of Students

Table 4.15: Teacher B with same number of Students

Cal-1 Teacher Z z Critical two-tail

A 4.008258 1.959964
C 3.370917 1.959964
D 5.288815 1.959964
F 4.463791 1.959964
G 4.119857 1.959964
H 2.997759 1.959964
E 3.630363 1.959964

Z-score data Analysis applied on the Cal-1 teacher “B” with all Teachers on the
basis different number of student. The value of Z-critical tow tail > Z conclude
that performance of teacher “E” is not significant with Teacher “A”, ,”C”, “D”,
“F”, “G”, “H”.,”E”
Table 4.15: Teacher B with different number of Students

Cal-1 Teacher Z z Critical two-tail

A 3.687091 1.959964
C 1.235961 1.959964
D 3.480026 1.959964
F 2.829936 1.959964
G 3.218075 1.959964
H 1.61421 1.959964
E 1.271677 1.959964

Z-score data Analysis applied on the Cal-1 teacher “B” with all Teachers on the
basis same number of student. The value of Z-critical tow tail > Z conclude
that performance of teacher “B” is not significant with Teacher “A”, “B”,”C”,
“D”, “F”, and significant with Teacher “G”, “H”. The ANOVA test has been
applied on the cal-I only. The results have been computed for the rest of the
courses similarly.

55
Pre-Qualification vs Programming Courses
In table 4.17, we have narrated the results of all programming courses which
are taught in 1st and 2nd semester. These averages of courses have been
compared with the Pre-Qualification.

Table 4.17: Comparison of all programming courses

ITC /ITP OOP/CP Simple

Teacher Pre-Q comb Simple Average Average
A 57.170543 61.48148148 65.27777778
B 61.4620758 53.7037037 62.80851064
C 60.99346693 65.21428571 69.5
D 61.04028149 61.81578947 72.39215686
E 60.28896102 71.40625 67.91666667
F 59.72643733 77 68.58333333
G 59.72643733 66.1871345 66.73484848
H 61.887987 66.9375 59.07142857
I 59.69578266 64.16129032 68.7826087
J 59.48424926 62.55882353 70.68421053
K 58.0787338 67.575 63.72222222
L 60.18534971 67.44736842 68.3125

Figure 4.9: Comparison of ITC/ITP teacher

According to the figure 4.8, the performance of all of the teachers has been

56
computed. The performance of teachers with respect to courses average
changes.

Average of ITC/ITP & CP/OOP Student

The average of students who studies ITC/ITP has been computed against every
teacher who teaches these subjects. Then these results have been shown in table
4.18 and presented in figure 4.10.

Table 4.18: Comparison of Actual and Predicted Grades of ITC/ITP

Pre-Qualification Predict ITC ITC /ITP

Teacher average /ITP average Simple average

A 57.170543 60.90678396 61.48148148

B 61.4620758 63.21581404 53.7037037

C 60.99346693 61.1117056 65.21428571

D 61.04028149 68.16813274 61.81578947

E 60.28896102 70.68016263 71.40625

F 59.72643733 69.13777079 77

G 59.72643733 63.65621403 66.1871345

H 61.887987 72.60489569 66.9375

I 59.69578266 65.2340331 64.16129032

J 59.48424926 70.26083656 62.55882353

K 58.0787338 61.80150053 67.575

L 60.18534971 65.17485658 67.44736842

57
 Figure 4.10: Comparison of Actual and Predicted Grades of ITC/ITP
Then the actual and predicted OOP/CP grades have been predicted against
every teacher. The results shows that there is a slight variation in the results
that has been stated in table 4.19 and presented in figure 4.11

Table 4.19: Comparison of Actual and Predicted Performance of OOP/CP

Teachers

Teacher OOP/CP ave Pred- OOP ave

A 65.27777778 65.99605885
B 62.80851064 58.81069353
C 69.5 67.8661911
D 72.39215686 66.23364694
E 67.91666667 72.16535104
F 68.58333333 73.23220667
G 66.73484848 67.4231563
H 59.07142857 65.26328164
I 68.7826087 67.14102048
J 70.68421053 67.37969774
K 63.72222222 66.97798275
L 68.3125 66.77153434

58
 Figure 4.11: Comparison of Actual and Predicted Performance of
OOP/CP Teachers
The overall performance of OOP teachers have been presented in table 4.20 as
well in figure 4.12.

Table 4.20: Overall Performance of OOP Teachers

Teacher Over All Performance

A 62.16653
B 60.00016
C 64.93713
D 65.93
E 68.49148
F 69.53595
G 64.74556
H 65.15302
I 65.00295
J 66.07356
K 63.63109
L 65.57832

The results show that the performance of teacher “E & F” is considerably high
and better than other teachers. Therefore, based on the acquired results, we may
recommend these two teachers to teach the courses in respective semesters.

59
 Figure 4.12: Overall performance of OOP Teachers

These results have been shown significant average of all students in ITC/ITP&
CP/OOP with ITC/ITP teacher name. During the computation of results of
ITC/ITP and CP/OOP, we have found teacher “F” performance better and able
to recommend teaching ITC/ITP and CP/OOP.

ANOVA data analysis on ITC/ITP

ANOVs analysis is used for group data we have many teacher for Cal-1 & Cal-
2 subjects. The average prediction shows that teacher average performance is
significant. To prove this result applies ANOVA analysis on this dataset.

If F > F crit, we reject the null hypothesis. This is the case 8.542703
>1.952212. Therefore, we reject the null hypothesis. The means of the 11
populations are not all equal. At least one of the means is different. Hence we
conclude that there are significant differences between the methods (i.e. all 11
methods don’t have the same mean).However, the ANOVA does not tell you
where the difference lies. Therefore it is preferred T-Test, Z-Score test each
pair of means.

60
 Table 4.21: ANOVA Data Analysis on ITC/ITP

Anova: Single Factor

SUMMARY

Groups Count Sum Average Variance

Row 1 6 370.458 61.743 9.09211

Row 2 6 346.3743 57.72906 16.59501

Row 3 6 404.7156 67.4526 4.601526

Row 4 6 397.287 66.21451 25.45386

Row 5 6 421.2417 70.20696 3.492003

Row 6 6 441.7835 73.63059 19.18218

Row 7 6 400.4038 66.73397 0.377307

Row 8 6 382.6283 63.77138 24.15625

Row 9 6 394.865 65.81084 6.082654

Row 10 6 392.0845 65.34741 12.9729

Row 11 6 382.7968 63.79946 15.06479

Row 12 6 407.9563 67.99271 0.222976

ANOVA

Source of
Variation SS df MS F P-value F crit

Between
Groups 1075.12 11 97.73818 8.542703 7.87E-09 1.952212

Within
Groups 686.4678 60 11.44113

Total 1761.588 71

Z-Score data analysis on ITC/ITP

As discussed earlier, Z-test is a measure to analyze the performance of two
attributes which have either equal or different number of population. Here, we

61
have applied Z-test on the combinations of teachers who teach ITC/ITP. We
have made combinations similarly as we made for the teachers of Cal-1 and
Cal-II. In table ..., Z-test has been applied on each pair of combination and
tested on two cases. First is to compare the performance of teachers on same
number of students and second is to compare the performance of teachers on
different number of students. Z-score data Analysis ITC/ITP has been applied
on the pair in which performance of teacher “F” with Teacher “A” on the basis
different number of students has been compared. The results show that there is
significant performance of teacher “F” with teacher “A”.

Table 4.22: Z-Score Data Analysis Teacher F with A different number of

Students

z-Test: Two Sample for Means

Variable 1 Variable 2

Mean 77 61.48148

Known Variance 260.2985 278.7682

Observations 67 27

Hypothesized Mean Difference 0

Z 4.116767

P(Z<=z) one-tail 1.92E-05

z Critical one-tail 1.644854

P(Z<=z) two-tail 3.84E-05

z Critical two-tail 1.959964

Z-score data Analysis has been applied on the pair of ITC/ITP teacher “F” with
Teacher “A” on the basis same number of students and results show that there
is a significant performance between teacher “F” and teacher “A”.

62
 Table 4.236: Z-Score Data Analysis Teacher F with A same number of
Students

z-Test: Two Sample for Means

Variable 1 Variable 2

Mean 76.81481 61.48148

Known Variance 283.7064 278.7682

Observations 27 27

Hypothesized Mean Difference 0

Z 3.359441

P(Z<=z) one-tail 0.000391

z Critical one-tail 1.644854

P(Z<=z) two-tail 0.000781

z Critical two-tail 1.959964

As computed earlier, Z-score data Analysis ITC/ITP teacher “F” with other
teachers who teaches same subjects on the basis different number of students
has been computed. According to the summarized results, there is significant
performance among teacher “F” with all of the teachers.

As computed earlier, Z-score data Analysis ITC/ITP teacher “F” with other
teachers who teaches same subjects on the basis of same number of students
has been computed. According to the summarized results, there is significant
performance among teacher “F” with all of the teachers.

63
Table 4.24: Z-Score Data Analysis Teacher F with other different number
of Students

Teacher F Z Z Critical two-tail

A 4.1168 1.96
B 7.0827 1.96
C 3.0686 1.96
D 5.4561 1.96
E 3.1293 1.96
G 4.7683 1.96
H 3.8317 1.96
I 3.8953 1.96
J 3.4076 1.96
K 3.3785 1.96
L 3.4421 1.96

Table 4.25: Z-Score Data Analysis Teacher F with other same number of
Students

Teacher F Z z Critical two-tail

A 3.3594 1.96
B 5.6999 1.96
C 2.4668 1.96
D 5.2634 1.96
E 1.323 1.96
G 3.9718 1.96
H 2.4788 1.96
I 3.1042 1.96
J 2.8519 1.96
K 2.6697 1.96
L 2.3868 1.96

64
Summary
To prove our point, whether student’s performance or grades are dependent on
teacher’s performance or not, we have taken an example to demonstrate. We
have selected the course cal-1 to justify our analysis. For this task, we have
taken courses; Cal-I & Cal-II and ITC/ITP & CP/OOP. Using different
approaches for on the basis these results, we have ended with following
assumptions.

There is no significant difference between actual and predicted average of

student’s grades on the basis of teacher’s combinations we discussed.

When Cal-1 and ITC/ITP is taught by different teacher and get average marks
of students from session 141 to 163 in Cal-1 and ITC/ITP teacher-wise using
five different experiments, like : Average of actual Cal-I marks based on the
Cal-I teacher, Prediction of the Cal-I marks excluding the teacher name and
then taking the average teacher-wise, Prediction of the Cal-I marks including
the teacher and then taking the average teacher-wise , Average of actual Cal-II
marks based on the Cal-I teacher , Prediction of Cal-II marks and then
computed the average of these marks based on the Cal-I teacher, our predicted
results were higher than the actual results or nearly lesser than actual ones. The
result has showed that all teachers have different average in these subjects and
performances of teachers in this subject are significant.

For proved these result have used ANOVA analysis .The performance of
teacher in CAL-I and ITC/ITP have different and average of teachers have
significant between all teacher .The teacher whose performance is better might
have a reason that he is lenient in giving grades to the students and the teachers
whose performance is comparatively low might be strict in giving grades to the
students. Z-Score have proved these results and have showed that significant
result in CAL-I and ITC/ITP according to each teacher. With the help of these
result, we have recommended for CAL-I teacher E is a very suitable teacher
and ITC/ITP teacher F can improve the performance of students.

We have proposed a recommender system, which will recommend teachers to

the students based on their interest level and understanding and communication

65
with their teachers. With the help of recommender system, university
department may become able to identify the student’s performance and enable
them to make their performance better. To evaluate our prediction model and
recommender system, we have produced the answers of research questions. We
have found following strong observations during the course of our experiments.

 In context of CUST university Cal-I and ITC are critical subjects in

semester-l

 Recommendation of appropriate teacher for Cal-l is B and G

 Recommendation of appropriate teacher for OOP is E and F

66
Chapter 5

CONCLUSION AND FUTURE WORK

 Prediction is a helpful activity to reduce the attrition

 Existing factor is applicable to the certain level of accuracy and a set of

attributes is identified which is applicable in context of our university

 In context of CUST university Cal-I and ITC are critical subjects in

semester-l

 Recommendation of appropriate teacher for Cal-l and ITC

Our proposed research has been distributed in five chapters. First chapter
consist of introduction of the topic and purpose of this research. Second chapter
contains literature review in which the related work of the other researchers has
been presently in quite understandable manner. Third chapter contains
methodology and chapter no 4 contains detailed discussion of results and
experiments.

In fist chapter, we have briefly explained the background of our research topic
along with its scope, significance and applications. Moreover, we have
discussed research question, which we have constructed after critically
reviewing of literature. We have explored various data mining techniques that
were proposed and used by many researchers. The area of data mining is quite
versatile. Data mining means to explore intrinsic information from the bulk of
datasets. These datasets might belong to educational institutions or internet or
any organization. With the help of mining techniques such as classification, we
can make data analysis for prediction and find the accuracy. Researchers have
worked a lot in the field of education in data mining in which they have used
student’s academic record for specific reasons. They have used the records for
grade prediction, evaluation of student’s performance and teacher’s
performance most commonly.

67
In this regard, from literature we have obtained exiting factors such as
institutional factors and demographic factors. These factors are known as
exiting because they have applied in the western universities. We have applied
these factors in our local context. We have collected data from the Capital
University of Science and Technology, Pakistan. We have used the data of
BSCS program from six terms of spring, which covers the data of 3 years. In
this dataset, we have used the data of students of first semester and their final
grade has been predicted based on their midterm marks, term marks, gender,
age, Matric grades and inter grade. These attributes cover demographic,
institutional and pre-university attributes of the students. In addition,
researchers have used only two factors; Demographic and Pre-university. In our
experiments, we have considered another factor which is institutional in which
midterm and term marks are combined to predict the students final GPA/Grade.
This is how we have proposed prediction model on the basis students
associated attributes.

After discussion of literature review critically, we have identified different

classification approaches by the researches, so we used in our experimental
work as well. Afterwards, we have discussed chapter number 3 in which
diagram of proposed methodology have been mentioned and each step of
methodology has been discussed in detail. Our chapter number 4 contains
relative experiments and results that we have carried out through the help of
WEKA tool. WEKA is specifically designed to perform analysis in the field of
educational data mining. In the part of conclusion we have summarize the
results against our each research question.

We have proposed a recommender system, which will recommend teachers to

the students based on their interest level and understanding and communication
with their teachers. With the help of recommender system, university
department may become able to identify the student’s performance and enable
them to make their performance better. To evaluate our prediction model and
recommender system, we have produced the answers of our three research
questions. Those research questions and their answers have been summarized
as follows:

68
RQ1. Whether exiting identified factors for grade prediction are valid in our
local context (Pakistan)?
The answer to this research question is yes because we have gained quite
considerable accuracy from all data analysis measures. These analyses have
been applied on categorical and nominal data. Results in both categories are
same, which shows the validity of, existing factors in local context. All three
factors contributed well in order to increase the accuracy through WEKA tool
when ARPP format of data was given to it and classification algorithms; SMO
and Linear regression were applied.

RQ2. Which factors help in accurate prediction of student’s GPA of first

semester?
We have given answer to this research question in quite interesting fashion. We
have make combinations of all attributes that belong to these three categories of
factors. When we combined term marks and mid-term grades along with inter
and matriculation marks as well as with their gender and age; accuracy of
algorithms were boost up to the mark able level. Therefore the answer to this
research question is that demographic, pre-university and institutional factors
(GPA, Internal marks, #of Credits Hours, Program). Based on these results, we
have proposed a prediction model that will help to predict the final GPA/Grade
of the students and then can be evaluated in other departments of the university
too. There is not a single attribute that is affecting the performance of students
but a combination of attributes that also occurs in terms of hybrid approach.

RQ3. Is it possible to improve the performance of a student by allocating of

appropriate teacher for a subject? To answer the research question no 3, we have
drawn inferences by two ways. One way is statistical experiments and other
way is prediction experiments. Both types of experiments have sub experiments
associated with them. by statistical experiments, we have computed ANOVA
test on the group of attributes to extract the information of such teachers having
significant performance. But ANOVA has a limitation; as it is applied in group
of attributes; therefore we are unable to identify the teacher having significant
performance. To overcome this shortcoming of inferences, we have applied Z-
test through which we have obtained the teachers having significant

69
performance with the pairing of each teacher belong to that particular course.
By this way, we became able to bring out the teachers whose performance is
better than other teachers who teach same course. Then we computed overall
average performance of the teacher based on the overall average grade of
students whom they taught.
With the help of prediction experiments, we have drawn prediction of grades
inclusively and exclusively teacher’s name who taught the courses. Then
predicted results and actual results have been computed and compared. By this
way we became able to get the entire set of results in which our analysis shows
that predicted and actual results remained close at some point and better at
some point too. In the same time, predicted were bit lower than actual ones as
well which can be seen in chapter 4.

At the end conclusion and future work has been discussed.

5.1 Future Work

● The experimental work can be applied in other departments of the
university.
● This data set is relatively small. We can carry out same experiments in
other field of department and large datasets.
Moreover, these experiments can be carried out in government institution such
as NADRA.

70
 REFERENCES

Abaya, S. A., & Gerardo, B. D. (2013, September). An education data

mining tool for marketing based on C4. 5 classification technique. In e-
Learning and e-Technologies in Education (ICEEE), 2013 Second
International Conference on (pp. 289-293). IEEE.

Aher, S. B., & Lobo, L. M. R. J. (2012). A comparative study of association

rule algorithms for course recommender system in e-learning. International
Journal of Computer Applications, 39(1), 48-52.

Al-Barrak, M. A., & Al-Razgan, M. (2016). Predicting student’s final GPA

using decision trees: a case study. International Journal of Information and
Education Technology, 6(7), 528.

Anand, S. S., Bell, D. A., & Hughes, J. G. (1996). EDM: A general

framework for data mining based on evidence theory. Data & Knowledge
Engineering, 18(3), 189-223.

Azarcon Jr, D. E., Gallardo, C. D., Anacin, C. G., & Velasco, E. (2014).
Attrition and retention in higher education institution: A conjoint analysis
of consumer behavior in higher education. Asia Pacific Journal of
Education, Arts and SCIENCE, 1(5), 107-118.

Baker, R. S., & Inventado, P. S. (2014). Educational data mining and

learning analytics. In learning analytics (pp. 61-75). Springer New York.
Baradwaj, B. K., & Pal, S. (2012). Mining educational data to analyze
students' performance. arXiv preprint arXiv: 1201.3417.

Bydžovská, H. (2013). Course Enrolment Recommender System (Doctoral

dissertation, Masarykova univerzita, Fakulta informatiky).

71
Cabrera, A. F., Nora, A., & Castaneda, M. B. (1993). College persistence:
Structural equations modeling test of an integrated model of student
retention. The journal of higher education, 64(2), 123-139.

Dougiamas, M., & Taylor, P. (2003). Moodle: Using learning communities

to create an open source course management system.

Drea, C. (2004). Student Attrition and Retention in Ontario's

Colleges. College Quarterly, 7(2), n2.

Elbadrawy, A., Polyzou, A., Ren, Z., Sweeney, M., Karypis, G., &
Rangwala, H. (2016). Predicting Student Performance Using Personalized
Analytics. Computer, 49(4), 61-69.

Garcia-Saiz, D., & Zorrilla, M. E. (2011, November). Comparing

classification methods for predicting distance students' performance. In
WAPA (pp. 26-32).

Guarín, C. E. L., Guzmán, E. L., & González, F. A. (2015). A model to

predict low academic performance at a specific enrollment using data
mining. IEEE Revista Iberoamericana de Tecnologias del
Aprendizaje, 10(3), 119-125.

Hung, J. L., & Zhang, K. (2008). Revealing online learning behaviors and
activity patterns and making predictions with data mining techniques in
online teaching. MERLOT Journal of Online Learning and Teaching.

Inoue, S., Rodgers, P. A., Tennant, A., & Spencer, N. (2017). Reducing
Information to Stimulate Design Imagination. In Design Computing and
Cognition'16 (pp. 3-21). Springer, Cham.

72
Kabakchieva, D. (2013). Predicting student performance by using data
mining methods for classification. Cybernetics and Information
Technologies, 13(1), 61-72.

Kaminski, J. (2005). Moodle–a User-Friendly, Open Source Course

Management System. Online Journal of Nursing Informatics, 9(1).
Kaur, G., & Singh, W. (2016). Prediction Of Student Performance Using
Weka Tool.

Kizilcec, R. F., Pérez-Sanagustín, M., & Maldonado, J. J. (2017). Self-

regulated learning strategies predict learner behavior and goal attainment in
Massive Open Online Courses. Computers & Education, 104, 18-33.

Kokina, J., Pachamanova, D., & Corbett, A. (2017). The role of data
visualization and analytics in performance management: Guiding
entrepreneurial growth decisions. Journal of Accounting Education.
Kovacic, Z. (2010). Early prediction of student success: Mining students'
enrolment data.

Merceron, A., & Yacef, K. (2005, May). Educational Data Mining: a Case
Study. In AIED (pp. 467-474).

Miller, B. N., Konstan, J. A., & Riedl, J. (2004). PocketLens: Toward a

personal recommender system. ACM Transactions on Information Systems
(TOIS), 22(3), 437-476.

Nam, C. S., & Smith-Jackson, T. L. (2007). Web-based learning

environment: A theory-based design process for development and
evaluation. Journal of information technology education, 6.

Oskouei, R. J., &Askari, M. (2014).Predicting Academic Performance with

Applying Data Mining Techniques (Generalizing the results of two

73
Different Case Studies). Computer Engineering and Applications
Journal, 3(2), 79-88.

Peña-Ayala, A. (2014). Educational data mining: A survey and a data

mining-based analysis of recent works. Expert systems with
applications, 41(4), 1432-1462.

Polyzou, A., & Karypis, G. (2016). Grade prediction with models specific
to students and courses. International Journal of Data Science and
Analytics, 2(3-4), 159-171.

Ramist, L. (1981). College student attrition and retention. Romero, C., &
Ventura, S. (2007). Educational data mining: A survey from 1995 to 2005.
Expert systems with applications, 33(1), 135-146.

Romero, C., & Ventura, S. (2010). Educational data mining: a review of the
state of the art. IEEE Transactions on Systems, Man, and Cybernetics, Part
C (Applications and Reviews), 40(6), 601-618.

Romero, C., & Ventura, S. (2013). Data mining in education. Wiley

Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 3(1),
12-27.

Sachin, R. B., & Vijay, M. S. (2012, January). A survey and future vision
of data mining in educational field. In Advanced Computing &
Communication Technologies (ACCT), 2012 Second International
Conference on (pp. 96-100). IEEE.

Scheuer, O., & McLaren, B. M. (2012). Educational data mining.

In Encyclopedia of the SCIENCE of Learning (pp. 1075-1079). Springer
US.

74
Tewari, A. S., Saroj, A., & Barman, A. G. (2015). e-Learning
Recommender System for Teachers using Opinion Mining. In Information
Science and Applications (pp. 1021-1029). Springer Berlin Heidelberg.

Tinto, V. (1987). Leaving college: Rethinking the causes and cures of

student attrition. University of Chicago Press, 5801 S. Ellis Avenue,
Chicago, IL 60637.

Yang, D., Sinha, T., Adamson, D., & Rosé, C. P. (2013, December). Turn
on, tune in, drop out: Anticipating student dropouts in massive open online
courses. In Proceedings of the 2013 NIPS Data-driven education
workshop(Vol. 11, p. 14).

Zaíane, O. R. (2002, December). Building a recommender agent for e-

learning systems. In Computers in Education, 2002. Proceedings.
International Conference on (pp. 55-59). IEEE.

Stallone, M. N. (2011). Factors associated with student attrition and

retention in an educational leadership doctoral program. Journal of College
Teaching & Learning (TLC), 1(6).

Romero, C., & Ventura, S. (2013). Data mining in education. Wiley

Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 3(1),
12-27.

Zhang, Y., Oussena, S., Clark, T., & Hyensook, K. (2010). Using data
mining to improve student retention in HE: a case study.

Witten, I. H., Frank, E., Trigg, L. E., Hall, M. A., Holmes, G., &
Cunningham, S. J. (1999). Weka: Practical machine learning tools and
techniques with Java implementations.

75
 Student Performance Prediction and
Teacher Recommender System

By

Muthara-Tul-Ain

A research thesis submitted to the Department of Computer

Science, Capital University of Science and Technology, Islamabad
in partial fulfillment of the requirements for the degree of

MASTER OF SCIENCE IN COMPUTER SCIENCE

DEPARTMENT OF COMPUTER SCIENCE

CAPITAL UNIVERSITY OF SCIENCE AND
TECHNOLOGY
ISLAMABAD
2017

i
 Copyright 2017 by CUST Student

All rights reserved. Reproduction in whole or in part in any form requires the
prior written permission of Muthara-Tul-Ain (MS133033) or designated
representative

ii
 CAPITAL UNIVERSITY OF SCIENCE & TECHNOLOGY
ISLAMABAD
Islamabad Expressway, Kahuta Road, Zone-V, Islamabad
Phone: +92 51 111 555 666, Fax: 92 51 4486705
Email: [email protected], Website: http”//www.cust.edu.pk

CERTIFICATE OF APPROVAL

Student Performance Prediction and Teacher Recommender

System
by

Muthara-Tul-Ain

MS133033

THESIS EXAMINING COMMITTEE

S No Examiner Name Organization

(a) External Examiner Dr. Ehsan Munir CIIT, Wah Cantt

(b) Internal Examiner Dr. Muhammad Arshad CUST, Islamabad

Islam

(c) Supervisor Dr. Nayyer Masood CUST, Islamabad

________________________________

Dr. Nayyer Masood

Thesis Supervisor
November, 2017
______________________________ ___________________________
Dr. Nayyer Masood Dr. Muhammad Abdul Qadir
Head Dean
Department of Computer Science Faculty of Computing
Dated : November, 2017
Dated : November, 2017

iii
 DEDICATED
TO

MY

RESPECTED

ALLAH (SWT) & PROPHET MUHAMMAD (PBUH)

&

PARENTS, BROTHER, SISTER

FOR

THEIR CONSTANT

AFFECTION AND SUPPORT

iv
ACKNOWLEDGMENT

All praise and exaltation is due to ALLAH (S.W.T) The creator and sustainer
of all seen and unseen worlds. First and foremost I would like to express my
gratitude and thanks giving to Him for providing me the boundaries and
blessings to complete this work. Secondly, I would like to express my sincerest
appreciation to my supervisor Dr Nayyar Masood for his directions,
assistance, and guidance. I sincerely thanked for his support, encouragement
and technical advice in the research area. I am heartily thankful to him from the
final level, as he enabled me to develop an understanding of the subject. He has
taught me, both consciously and unconsciously, how good experimental work
is carried out. Sir you will always be remembered in my prayers.

I am highly indebted to my parents and my family, for their expectations,

assistance, support and encouragement throughout the completion of this
Master of Science degree. They form the most important part of my life. After
ALLAH (S.W.T) they are the sole source of my being in this world. No words
can ever be sufficient for the gratitude I have for my parents and for my family.
A special thanks to my employing company for their support and
encouragement to complete this Master of Science degree.

I pray to ALLAH (S.W.T) that may He bestow me with true success in all
fields in both worlds and shower His blessed knowledge upon me for the
betterment of all Muslims and whole Mankind.

AAMEEN

Muthara-Tul-Ain

v
DECLARATION
It is declared that this is an original piece of my own work, except where
otherwise acknowledged in text and references. This work has not been
submitted in any form for another degree or diploma at any university or other
institution for tertiary education and shall not be submitted by me in future for
obtaining any degree from this or any other University or Institution.

Muthara-Tul-Ain

Reg. No, MS133033

November, 2017

vi
 ABSTRACT
Mining data and extracting information from huge databases has become an
interesting research area for the researchers. The idea to extract information
with the help of data mining techniques came into being since a couple of
decades ago. Initially, researchers were supposed to apply classification and
clustering techniques to partite the dataset and analyze the intrinsic features. On
the basis of such features, they make reasonable predictions. These predictions
have taken place in the field of educational data mining for many purposes such
as; predict the performance of students on the basis of factors associated with
them, to enable them suitable courses and appropriate teachers. These purposes
have been derived from the area of student retention and attrition. Our research
aims to achieve these purposes under the roof of student attrition and retention.
Moreover, we have identified such exiting factors which are beneficial for
predicting the performance to students, recommend them best suitable teachers
and help them to select the courses. We have applied classification algorithms
with respect to the nature of data which have been collected from the Capital
University of Science and Technology (CUST), ISB. In this study, the GPA of
first semester on the basis of Midterm and previous academic grades have been
tried to predict. For the second semester, we have predicted the CGPA of same
students by using their complete proceeding academic record with the help of
hybrid approach. The hybrid approach consists of combination of factors that
have been evaluated against our research questions. Moreover, we tried to
improve their performance by recommending those suitable courses and
teachers whose performance is better amongst others comparatively. The
reason of collecting the data from CUST is to validate the exiting factors in
local context (Pakistan). On the basis of classification algorithms such as;
Naïve Bayes and J48, we have become able to build the recommender system.
Then the factors which contributed well to validate the exiting factors in local
context have been measured. In the last, appropriate teacher allocation has been
measured by two ways; Statistical and Prediction. In statistical
experimentations, average performance of teachers, Z- test and ANOVA test
has been applied. In prediction experimentations, one subject teacher with other

vii
subject teacher’s name attribute, and one subject teacher without other subject
teacher’s name attribute and overall performance of teachers have been
computed with respect to each subject independently. With the help of our
research we might have become able to provide a way to educational
institutions to reduce the attrition and increase the retention rate.

viii
 Table of Contents
Abstract .........................................................................................................................vii
Introduction ..................................................................................................................... 1
1.1 Background Of Research ......................................................................................... 2
1.2 Problem Statement ..................................................................................................... 6
1.3 Research Questions.................................................................................................... 6
1.4 Scope ......................................................................................................................... 6
1.5 Application Of Proposed Approach............................................................................ 6
1.6 Significance Of Solution ............................................................................................ 7
1.7 Organization Of Thesis .............................................................................................. 8
1.8 Definitions And Frequently Used Terms .................................................................... 8
Literature Review ............................................................................................................ 9
2.1 Educational Data Mining ......................................................................................... 10
2.2 Student Performance Prediction ............................................................................... 14
2.3 Commonly Used Approaches In Edm ...................................................................... 16
2.5 Commonly Used Attributes For Student Performance Prediction. ............................ 19
2.5 Literature Review Summary .................................................................................... 21
Research Methodology .................................................................................................. 22
3.1 Data Collection And Pre-Processing ........................................................................ 24
3.2 Classification ........................................................................................................... 27
3.3 Algorithms And Techniques .................................................................................... 29
3.4 Evaluation Of Research Questions ........................................................................... 30
Results & Evaluations.................................................................................................... 35
4.2 Final Attribute Selection .......................................................................................... 38
4.3 Result Of Evaluation Of Research Questions ........................................................... 38
Conclusion And Future Work ........................................................................................ 67
5.1 Future Work ............................................................................................................ 70
References ..................................................................................................................... 71

ix
 List of Tables
Table 2.1: Literature Review Summary ........................................................................................... 21
Table 3.2: Summary Of Session-Wise Attrition............................................................................. 24
Table 3.3: Summary Of Session-Wise Attrition............................................................................. 25
Table 3.4: Selection Of Attributes ..................................................................................................... 28
Table 3.5: Attributes Selection To Dataset ...................................................................................... 31
Table 3.6: Pattern Of Combination Of Attributes .......................................................................... 32
Table 4.1: Occurrence Of Missing Values ....................................................................................... 36
Table 4.2: Handling Of Missing Values ........................................................................................... 37
Table 4.3: Attributes Table ................................................................................................................... 39
Table 4.4: Classification Of Attributes And Algorithms ............................................................. 40
Table 4.5 Ranked Attributes ................................................................................................................. 42
Table 4.6 GPA PREDICTION FOR THE FIRST SEMESTER .............................................................................. 42
Table 4.7: Cgpa Prediction For The Second Semester ................................................................. 44
Table 4.8: Pre-Qualification Vs Maths Courses ............................................................................. 47
Table 4.9: Average Of Cal-Ll Teachers ............................................................................................ 48
Table 4.10: Average Of Cal-I Teachers ............................................................................................ 49
Table 4.11: Overall Average Of Maths Teachers ......................................................................... 50
Table 4.12: Anova Analysis On Cal-1 .............................................................................................. 52
Table 4.13: Two Samples For Mean With Respect To Teacher “E”........................................ 53
Table 4.14: Two Samples For Means With Respect To Teacher “B” ..................................... 54
Table 4.15: Teacher B With Same Number Of Students ............................................................. 55
Table 4.16: Teacher B With Different Number Of Students ...................................................... 55
Table 4.17: Comparison Of All Programming Courses ............................................................... 56
Table 4.18: Comparison Of Actual And Predicted Grades Of Itc/Itp ...................................... 57
Table 4.19: Comparison Of Actual And Predicted Performance Of Oop/Cp
Teachers ............................................................................................................................................ 58
Table 4.20: Overall Performance Of Oop Teachers ...................................................................... 59

x
 List of Figures
Figure 3.1: DATA FLOW DIAGRAM OF PROPOSED METHODOLOGY ......................................................... 23
Figure 4.1: Formats Of Data................................................................................................................. 36
Figure 4.2comparison Of Local And Existing Attributes ............................................................ 41
Figure 4.3: Gpa Prediction For The 1st Semester .......................................................................... 44
Figure 4.4: Cgpa Prediction For The Second Semester................................................................ 45
Figure 4.5: Pre-Qualification Vs Maths Courses ........................................................................... 48
Figure 4.6: Average Of Cal-Ll Teachers .......................................................................................... 49
Figure 4.7: Overall Average Of Maths Teachers ........................................................................... 50
Figure 4.8: Comparison Of Itc/Itp Teacher ...................................................................................... 51
Figure 4.9: Comparison Of Actual And Predicted Grades Of Itc/Itp ....................................... 56
Figure 4.10: Comparison Of Actual And Predicted Performance Of Oop/Cp
Teachers ............................................................................................................................................ 58
Figure 4.11: Overall Performance Of Oop Teachers .................................................................... 59
Figure 4.12: Overall Performance Of Oop Teachers .................................................................... 60

xi
 List of Abbreviations
CUST: Capital University of Science and Technology

WEKA: Waikato Environment for Knowledge Analysis

MOOCs: Massive Open Online Courses

CSV: Comma-separated values (Comma-delimited)

ARFF: Attribute-Relation File Format.

GPA: Grade point average

CGPA: Cumulative Grade Point Average

Cal-I: Calculus-I

Cal-II: Calculus-II

ITC: Information Technology Computing

ITP: Information Technology Programming

CP: Computer Programming

OOP: Object Oriented Programming

SEM: Semester

xii
Chapter 1

INTRODUCTION
Researchers have been working in the area of educational data mining for a
decade. Educational Data Mining (EDM) has become a broadened field in
which researchers are conducting their experiments to extract useful
information from the data belonging to the educational sectors for many
purposes. The purposes include identifying student attrition and retention rate,
students’ performance prediction, building a course recommender system, and
teacher recommender system, etc. In the field of Educational Data Mining
(EDM), researchers are busy in exploring effectiveness or role of different
types of variables to measure and predict the performance of students. Among
those variables, student’s age, academic record and biography of students are
involved. EDM is a growing research field, which carries data mining
techniques in educational system (Romero, C., & Ventura, S. 2007). In
previous study, (Romero, C., & Ventura, S. 2013) has explored the
phenomenon of making the student’s outcome better with the help of data
mining approaches. According to them, huge data from the institutions have
problems associated with it. Therefore, data mining approaches are not
supposed to be applied directly on this data. Therefore, knowledge discovery
process has been implemented. In educational institutions, data mining is
playing a pivotal role on datasets after preprocessing. Most widely used data
mining approaches involved classification and clustering, outlier detection,
association rule mining, pattern mining and text mining. Researchers are trying
to extract useful, novel and interesting information with the help of data mining
approaches (Romero, C., & Ventura, S. 2010). EDM process is used to convert
raw educational data into useful information.

In the field of EDM, researchers are active in three broad categories; Personal
recommender system and learning environment, course management system,
and student attrition and retention. Our research will cover all these mentioned
areas in different contexts such as identification of students with low academic
performance, building prediction model to predict the student’s performance by

1
using their historical data, improvement in raising the confidence level among
them, assist them to choose their courses, building a model that will assign
appropriate teachers to the students.

1.1 Background of research

As discussed in above section, EDM is subfield of data mining which is applied
in educational institutions to help them to maintain their prestige. Educational
data mining is all related to propose and develop naval approaches to improve
the performance of students on a large scale. Educational data mining is an
active research areas in which researchers are working actively in three main
areas of educational data mining are:

a) Personal recommender system and learning environment

Researchers are using this concept in hybrid fashion in the context of
educational data mining. They are constantly using various data mining
approaches to make the personal recommender system better to produce
maximum accurate results. In a research (Miller, B. 2004), personal
recommender system uses collaborative filtering to reduce noise from the data.
A limitation of personal recommender system is found to be its un-portability
as they are operational only on large computers that are connected to the
internet. Another limitation was found to be its trustworthy relationship
between the owner of recommender system and its user.

b) Course management system and educational data mining

Researchers are associated with this branch of EDM to provide the universities
and educational institutions a platform that can help them in management of
courses in the department. For this purpose, different data mining approaches
have been carried out (Peña-Ayala, A. 2014). The concept has been derived
from e-learning system “Moodle”. Moodle is a learning management system
that helps educators to build up effective online learning communications
through the analysis of information generated by the users (Aher, S. B, 2012).

2
c) Student attrition and retention
Student attrition means the number of students who are leave their courses
without completing because of certain reasons. Those reasons may include poor
choices of courses, consistent undesirable result, poor performance, financially
instability and immature selection of courses etc whereas student retention
means the number of students who complete their course and acquire degree
despite of any kind of circumstance and gain good grades in the transcript. In
educational institutions, rate of student attrition and retention has reached up to
considerable unit (Stallone, M. N. 2011).

Student attrition and retention not only affect the performance of education
institution or department but also affect the faculty positions and raise the
financial problems for the parents eventually. To analyze the pattern of student
retention and attrition, one has to find out the reasons of this cause then take
steps to resolve this problem. Our research topic falls into the type student
attrition and retention which is further explained in the following subsection in
detail. For this research collected dataset from Capital University of science
and Technology, Department of Computer Science. Let’s consider first
semester and second semester courses because student attrition rate is high in
first and second semester. These semesters are helpful attrition of students and
predicted the student performance.

Student performance prediction

To recognize the efforts of students in academia, there are some known
parameters carried out by universities or educational institutions (Romero, C.,
& Ventura, S, 2007). Among those parameters, student’s grades, attendance,
grades per courses are commonly evaluated. Student performance prediction
has been the growing research problem for the researchers since a decade. To
improve the institution’s prestige and efficiency with respect to student
performance improvement, there are several techniques used by different
researchers such as course specific regression, personalized linear multi
regression, classification and clustering. These approaches not only work for
educational institutions those are in existence physically but also for virtual
campuses too i-e; LMS and Moore of Stanford University (Elbadrawy, et al,

3
2016). EDM is a broad research field in which researchers are exploring a lot of
problems including student’s performance, attrition and retention, course
selection, and particular teacher’s selection. The widely covered research
problems in the domain of educational data mining are known as student’s
future grade prediction, performance prediction and course enrollment
recommender system.

This research is based on different aspects. One aspect is to identify the number
of those students who had to leave the institution because of their poor
performance in studies or financial problem. Then the number of those students
have been found who completed their degree belong to the spring semesters of
three years. Expectedly this research will help the students to improve their
results in the courses which have been taken for the experiments by
recommending them appropriate teachers and courses based on their associated
factors.

This is helpful not only for students to raise their grades but also for parents
not to suffer from financial problems. Another aspect which is being
considered is to retain maximum percentage of students in the university and
fulfill their degree requirements till its completion. Ultimately, if the grades of
students would be good their retention will become stronger in the institution.
This research might reduce the attrition rate because of the provision of teacher
recommender system, course recommender system and student performance
prediction. The outcome of our research has been discussed in detail as follows.

Students Performance Prediction is the research problem which has been

identified by many researchers with the help of different approaches such as
Matrix Factorization, classification and clustering algorithms (Polyzou, A., &
Karypis, G, 2016). Before applying these approaches, students variables are
supposed to be extracted from the collected dataset that are unique registration
number, result of previous two terms, GPA (Grade Point Average), number of
family members, family welfare, gender, usage of social networking and use of
technology and internet (Oskouei, R. J., & Askari, M, 2014). After collection,
data preprocessing is applied in which positive and negative effects of dataset

4
are removed and then classification and clustering are applied. By using such
techniques, data can be further divided into clusters and classes to analyze
useful and hidden patterns from it (Kabakchieva, D, 2013).

With respect to performance prediction of students, there are many cases such
as student’s next term grade prediction, presence performance based on
assessment in the current courses. To evaluate this problem, researchers have
used some of the characteristics of students such as admission records, High
school scores, SAT/ACT scores and grades of previously completed courses
(Elbadrawy, et al, 2016). More characteristics such as class test, seminar
attendance and marks, assignment marks (Baradwaj, B. K., & Pal, S. 2012).

Course enrollment recommender system It helps to determine the

performance of students by recommending them courses up to their
performance. For this purpose researchers have proposed the recommender `in
which he used 67 templates of active students. This recommender system not
only helps to predict the grades of students, it also helps to recommend the
courses to the students by considering their timetable (Bydžovská, H, 2013).

Proposed attributes of students affects their performance in various ways. It not

only helps in identifying the students with low academic performance but also
paves the ways to improve their grades and enables recommender systems of
universities to predict the grades of students. For this purpose, different
classification techniques of data mining have been studied in this research.
With the help of this study, student carrier can be affected by means of building
up and students can easily pave their way towards their desired destination.
This study will also help to construct the course recommender system which
will allow the students to select their courses that are recommended by the
proposed system. Another advantage of this study is highly beneficial
economically for parents. With the help of proposed system, students may not
waste their money in selection of courses to repeat. Moreover, student’s
academic record can be improved and efficiency of department management
can be raised.

5
The factors used in the research have already been applied in countries other
than Pakistan. Those countries are Canada, USA, England, and Nigeria. All
these factors have been applied in the colleges and universities of narrated
countries and now will be validated in the context of Pakistan.

1.2 Problem Statement

A major problem that the academic institutions are facing worldwide is poor
performance of students in academics. This causes high attrition rate that is a
loss to students, parents and institutions. Student’s performance prediction can
help to reduce attrition rate as it raises an early alarm for students not
performing well and are likely to leave the institution/studies. Most of the work
on performance prediction has been done in foreign institutions; this has to be
done in the context of Pakistan for more accurate results.

1.3 Research Questions

Given below is the proposed problem statement

RQ1. Whether exiting identified factors for grade prediction are valid in our
local context (Pakistan)?
RQ2. Which factors help in accurate prediction of students’ GPA of first
semester?
RQ3. Is it possible to improve the performance of a student by allocating of
appropriate teacher for a subject?

1.4 Scope
The application of this research is versatile. The experiments of this research
have been conducted on the dataset that has been collected from Capital
University of Science and Technology, Islamabad (Pakistan). Such factors can
be applied on the data of Government institutions as well as schools and
colleges.

1.5 Application of Proposed Approach

Our proposed research will be beneficial in one of the following ways in the
domain of educational data mining.

6
To make the university management system efficient
This research might be highly beneficial for the university management system
to keep check on the performance of student’s belong to different departments.
This will not only help to predict and improve student’s performance but also
supports the parents economically.

To enable Course recommender system pro-active

The selection of courses is major problem for the students. To build such
recommender system that helps students in selection and registration of the
courses will be quite beneficial for the institutions.

Helpful in improving teacher recommender system

In the proposed solution this recommender system will help department to
recommend specific teachers to particular students based on previous 0239
academic record and inclination of selected courses in which they pursued good
grades.

1.6 Significance of solution

This research is a fine contribution for the educational institutions of Pakistan.
With the help of this research, the factors that have been used by the
researchers from the foreign universities are considered in local context. These
factors are planned to apply over the data set of Capital University of Science
and Technology of Pakistan. These factors have been utilized with the help of
WEKA for the student’s performance prediction. Universities and other
educational institutions may capable of predicting the student’s next term
grades.

Moreover, with the help of the contribution of this research there are chances to
propose teacher recommender system that will help to recommend suitable
teachers to the students by considering their current performance in the
semester and their relative interest in the courses that might improve their
grades.

7
1.7 Organization of thesis
This thesis comprises of five chapters. The first chapter states the introduction
of the proposed research. Second chapter is the literature review in which the
work and research contribution in chosen area by former researchers has been
discussed. Moreover, overall literature summary has been presented in
summarized way. Third chapter contains the methodology diagram that has
been adopted to perform the experiments and how experiments will be
performed. The results of experiments by using selected tools have been
presented in the chapter four. Last chapter contains the conclusion and future
work in which summary of whole thesis has been presented.

1.8 Definitions and frequently used terms

Performance Prediction
A student’s grade is used to acknowledge his/her related performance in
academia. It is now possible to predict the student’s performance with scientific
methods in which quantitative approaches have been used by the researchers
(Bhardwaj, B. K, 2012) (Pal, S. 2012).

Recommender System
Usually recommender system is built to assign a particular entity to required
entity. In the field of data mining, researchers have proposed some
recommender systems in which teacher recommender system and course
recommender systems are involved (Bozo, J., 2010) (Alarcón, R.,2010)(
Iribarra, S., 2010).

8
Chapter 2

LITERATURE REVIEW
The initiative of online course learning system is based on e-commerce
venture. With the growth of this venture, data from web resources started
collecting and storing in the excel that contains customer and product
information and order information (Kokina, J, 2017). E-commerce is a term
that refers to use as online business through internet. There are various websites
that are working for this purpose such as e-bay, Alibaba, and Amazon etc. It
can be said that data storage in excel from the resources has been derived from
e-commerce. Predicting student’s performance by using data mining techniques
to extract information from the academic dataset of universities has become
state of the art research in the scientific society. Universities are confronting
with some challenges now a day to analyze the performance of their students.
That’s why researchers are focusing on student’s profiles and characteristics to
make the university management aware of student’s performance and overall
academic result (Kabakchieva, D, 2013). There is another dimension of
student’s performance that is the dependence of student retention upon student
student’s performance. To minimize the problem of student retention cases in
the universities, different researchers have proposed different methods to
predict the performance of students in their future semester based on the
performance of previous one.

To predict the courses of next term grades, four parameters have been
considered in this study such as; admission records, High school scores,
SAT/ACT scores and grades of previously completed courses. Based upon
these parameters, recommender system can be trained to predict the grades of
students accurately in any of the educational institution. Historical information
about the course has also been considered in this study such as which course is
taught by which teacher and information about contents of the course. Many
researchers have used LMS and Moore to predict the successive chances of
success and failure of students. In this research, regression based methods such
as course specific regression (CSPR) and personalized linear multi regression

9
(PLMR) has been used. Another method known as matrix factorization based
methods in which standard matrix factorization (MF) has been used for the
grade prediction of students (Elbadrawy, A, et al, 2015).
In this chapter, the background of educational data mining and its branches will
be discussed in detail. Also the student’s performance prediction and data
mining approaches that are commonly used by researchers in the literature are
being discussed.

2.1 Educational Data Mining

Data mining is a process of sorting data and extracting information from
existing databases.1 With the help of pattern mining and data analysis, hidden
information can be obtained from huge datasets. The strategy of data mining is
now applied in the field of education by researchers. They are busy in
exploiting a lot of dimensions in education sector. This is now known as
educational data mining. Data mining is applying in educational sector by
considering the performance of students and finding the position of students by
using their academic records. Educational dataset is being collected from
various resources such as interactive learning systems, computer-supported
collaborative systems, and administrative datasets of school, colleges and
universities2. Data mining methods are now implemented in well known
universities to analyze the patterns of student performance from the dataset
through which information can be extract and decision making may become
easier for the management of institutions (Kabakchieva, D. 2013).

With the incremental growth in the use of technology everywhere, educational

institutions are now busy in finding hidden trends and patterns in their larger
datasets (Scheuer, O. 2012). (Merceron, A. 2005). Datasets that are used for
experimental purposes in educational institutions have become possibly
available because of web based educational systems in which LMS, Moore and
Portal system have become common. With the help of these sources, dataset
can easily be collected if authorization is accessed. One purpose of extracting
information from its own dataset is to make its prestige among other

1
http://searchsqlserver.techtarget.com/definition/data-mining
2
http://www.educationaldatamining.org/

10
educational institutions stronger. Another purpose is to build the student career
by covering its each and every aspect such as improvement of their grades if
lacking, overall performance booster, support them financially, make them
enable to select the courses suggested by course recommender system, assign
them appropriate teachers based on their inclination of interest in course
selection and many more.

However, these are some common areas in which researchers are producing
their research by using different data mining techniques. In our research, same
areas are being covered under the dataset of Pakistani students. Three factors
are being applied on students those factors have been discussed briefly in
section 2.4. In current section, our focus will remain on key areas of EDM, in
which researchers are engaged.

a) Personal recommender and learning environment

Web based learning environments have become common right now. It has
become powerful medium to act as a bridge between learner and instructor and
provides interesting learning mechanisms to both (Nam, C. S, 2007) (Osmar,
R, 2002) has elaborated the concept and working of web based learning which
is also known as e-learning. A recommender system helps to recommend the
actions to a learner by using his previous actions with the help of intelligent
software agents. The use of such recommender systems were first initiated in e-
commerce and now carrying out by researchers in the domain of e-learning.
Osmar has implemented this recommender system in assisting the offline web
miners who are responsible to find hidden trends in the data and online course
navigation. With the help of recommender system, educational institutions are
now strengthening their management and departmental progress. Researchers
have proposed several recommender systems among which course
recommender system and teacher recommender system are commonly known.

b) Course management system

By means of course management, researchers have explored different sources
to find the best possible use of course management system. In this context,
Moodle is found to be open source course management system that helps

11
educators to make their courses available on the internet. All the data on the
Moodle is managed by Moodle team. Moodle is now facilitating educational
institutions, community colleges, and schools to create online teaching system
(Dougiamas, M., 2003). It delivers courses online, unlike traditional
classrooms. Data of courses and students is stored in its specified database
which is further used by researchers that carries out for mining purposes to
extract useful information from it. Performance of Moodle has raise higher in
virtual environment. Other online course management system such as web
portals and learning management systems are serving for the same purpose now
days.

With the use of internet in classrooms and institutions, those institutions have
launched their proper websites or web pages, through which students can
register themselves into particular courses (Kaminski, J. 2005). In the area of
online course learning, researchers have presented many research articles. In
this domain LMS and Moore have contributed a lot. They have been offering
different courses on their websites and thousands of students from all over the
world registered themselves. This is how, huge amount of dataset is collected
from LMS and Moore and researchers have performed analysis on these dataset
to evaluate the performance of students who enrolled (Kizilcec, R. F, 2017)
(Pérez-Sanagustín, M., 2017) (Maldonado, J. J. 2017). Self regulated learning
(SRL) is a term that has been stated by narrated researchers. According to
them, students having strong SRL are good enough in planning, managing and
controlling as compared to the students having weak SRL. In this regard,
MOOCs have been providing support for the learners with the help of different
levels of SRL.

c) Student attrition and retention

With the passage of time, growth of private educational institutions has been
increased up to the remarkable extend. These institutions have become source
of higher learning and business entity. Therefore, maximum number of
student’s enrollment is its lifeline. For the survival of private institutions,
profitability, proper management and alignment are mandatory. In this respect,
student retention until the completion of degree is quite necessary. That’s why

12
institutions are finding that factor that ultimately causes student attrition
(Azarcon Jr et al, 2014). After analyzing those factors, it is important for
educational institutions to make strategic adjustments accordingly to improve
student retention in institutions. In Gaviria, Colombia, people are closely
belongs to social mobility that is a causing the academic performance of the
students and student attrition rate may increase in educational institutions.
Researchers have used three drivers the student performance analytics. The first
driver is the volume of data that is being collected from learning management
system and student information system, second one is the e-learning and third
one is political concerns (Guarín, C. E. L., 2015)( Guzmán, E. L.,2015)(
González, F. A. 2015). The application of data mining in the field of
educational data mining has been emerged in different areas and researchers are
exploiting these areas with various dimensions.

There is another supplement for student learning through web that is known as
MOOCs. In online course learning and management system, high rate of
dropout of students have been identified by researchers. This problem has been
enlightened by the YANG in his research that is students are dropping at
considerable level in Massive Open Online Courses (MOOCs). To control the
student attrition problem from the coarser classes, researchers have proposed a
model in their research. This model is a helping hand to determine such
influential factors that are causing student drop out. So the predictors have been
tried to propose in this research that will determine the factors related to
student’s behavior and social position in the discussion in which they
participate in the forums (Yang, D.et al, 2013).

The problem of student attrition and retention is not new for the educational
institutions. It has been enlightened by the researchers from the fields of data
mining and information visualization. Now it has become very common
research problem for the researchers. Student attrition and retention problem
has been observed by the researchers when this problem was raised up to the
ratio of 50% on the colleges of Ontario (Drea, C. 2004). To reduce attrition
rates, institutions should focus on student retention. Researchers have analyzed
the factors that causes student attrition and in the research, Drea has addressed

13
both elements; student attrition and retention. To retain the persistence of
institutions, two theories have been formulated in this context. The first one is
student integration model by Tinto and student attrition model by Bean. Both
models work almost similar (Cabrera, A. F. 1993).

There are numerous reasons for what student attrition has become research
problem for the researchers. In those reasons, personal disappointments,
financial setbacks, and lowering of career and life goals are considerable.
Therefore scientists have carried out the retention and persistent in their
research to resolve this societal problem (Ramist, L.1981). This research focus
on student performance and also find a teacher methodology has positive
impact on student grade prediction and has reduced student retention rate.

2.2 Student performance prediction

Student performance prediction is the research problem which has been
considered by many researchers from the field of educational data mining
(Kabakchieva, D, 2013). With respect to performance prediction of students,
there are many cases such as student’s next term grade prediction, presence
performance based on assessment in the current courses. To evaluate this
problem, researchers have used some of the characteristics of students such as
admission records,

High school scores, SAT/ACT scores and grades of previously completed

courses (Elbadrawy, et al, 2016). More characteristics such as class test,
seminar attendance and marks, assignment marks (Baradwaj, B. K., & Pal, S.
2012). For the sake of performance prediction, data that has been stored in
open polytechnic student management system belong to 2006-2009 had been
used. For the experiments to be applied over the dataset, feature selection and
clustering techniques were applied on the dataset which are also known as data
mining techniques.

In the field of education, predicting the performance of students has become

considerably important for the university management and departments.
Because the prestige of any educational institution is ultimately depends upon
the performance of its students and if performance of maximum number of

14
students degrade up to the mark able level, it should be considered by the
department management and find out the factors that affects the performance of
students. Under consideration of this problem, researchers have evaluated their
studies against the parameters that affects the overall performance of students,
data mining techniques and data mining tools (Kaur, G, 2016)( Singh, W,
2016). In such parameters, psychological, personal and environmental factors
are involved. For the experiments, they have used Naïve Based and J48
techniques through WEKA.

a) Future grades prediction

To predict the future grades of students is the considerable research problem

that has been identified by many researchers with the help of different
approaches such as Matrix Factorization, classification and clustering
algorithms (Polyzou, A., & Karypis, G, 2016). Before applying these
approaches, students variables are supposed to be extracted from the collected
dataset that are unique registration number, result of previous two terms, GPI
(Grade Point Average), number of family members, family welfare, gender,
usage of social networking and use of technology and internet (Oskouei, R. J.,
& Askari, M, 2014). After collection, data preprocessing is applied in which
positive and negative effects of dataset are removed and then classification and
clustering are applied. By using such techniques, data can be further divided
into clusters and classes to analyze useful and hidden patterns from it. To
predict the performance of students and predict their grades, researchers have
used socio-demographic factors of the students such as age, gender, ethnicity,
education, work, status, and disability. In addition, researchers have used
environment may also affect the student’s performance that causes their
dropout or retention throughout the course accomplishment (Kovacic, Z. 2010).

b) Teacher recommender system

When students retain their degree the end of their duration, it encompasses the
quality education of any educational institution that refers to the term of student
retention. According to the national statistics in a research, fifty percent of the
students from higher educational institutions get attired. To (Bozo, J., 2010)

15
(Alarcón, R., 2010) (Iribarra, S., 2010). As online course enrollment system has
taken the place globally because of internet. In this variation in technology,
there are several web pages and websites that are busy in providing best at their
own but still contain some missing elements in their services. Therefore,
researchers have worked upon this problem and proposed solution regardingly.
To make every course learnable creatively, researchers has focused on teacher
recommendation according to the course respectively. The new recommender
system is known as A3. With the help of this recommender system, not only
accurate teacher will be associated to the relative course contents but also best
contents of the course will be available on that site after some time (Tewari, A.
S, 2015).

c) Course recommender system

A course recommender system helps to determine the performance of students

by recommending them courses up to their performance. For this purpose
researchers have proposed the recommender in which he used 67 templates of
active students. This recommender system not only helps to predict the grades
of students, it also helps to recommend the courses to the students by
considering their timetable (Bydžovská, H, 2013).

2.3 Commonly used approaches in EDM

From the literature, a lot of data mining approaches that have been carried out
by the researchers to evaluate their studies. Most commonly used approaches
have been discussed in detail in subsections of section.

a) Classification and Clustering

To improve the performance of weak students in the class, with the help of data
mining techniques, collected data of 1100 students have been transformed in
Weka. Researchers have used freeware software such as Weka, Clementine and
Rapid-Miner in this work. Different classifiers such as Naïve Based, C4.5,
Neural networks and random forest have been used. The classifiers which have
been used are Adaboost, Bagging and boosting. It is found that some factors
have same effects on both countries and some have different. Moreover, it has

16
been found that male students suffer with stress more as compared to female.
The performance of male students in Mathematics and formal SCIENCE is
better whereas performance of female students aroused better in Literature and
mnemonic SCIENCE (Oskouei, R. J.,2014)( Askari, M. 2014).In this field of
research, there exist wide varieties of benchmark which are used to evaluate the
performance and accuracy of experiments conducted by using machine learning
approaches. Different researchers have used various types of educational
datasets and each dataset is unique in its attributes (Garcia-Saiz, D., 2011)
(Zorrilla, M. E, 2011). Therefore, in the study researchers have proposed meta
algorithm to preprocess dataset. Various data mining models have been studied
in this research to find the most accurate one with the help of Meta algorithm.

Educational Data Mining approach is emerging in providing feasibility to the

educational institutions to improve their teaching and learning methodologies.
EDM is the process of applying data mining techniques in educational
institutions. In educational institutions, the data of student’s performance,
teacher’s evaluation, student’s enrollment data, and gender differences are to be
stored in the database. On this data, Classification, Clustering and association
rules are applied. These rules have been applied on mandatory courses taught in
IT department of King Saud University. To extend this approach, these rules
can be applied on elective courses as well to predict the grades of students in all
courses as well as final GPA. Furthermore, neural network and clustering can
also make these rules effective for prediction (Al-Barrak, M. A., & Al-Razgan,
M. 2016). Researchers have considered more attributes of students such as their
attendance, class test, seminar attendance and marks, assignment marks and
constructed classification tree in Weka. According to those researchers,
teachers can minimize the failure ratio of students in the semester by adopting
this approach and attributes (Baradwaj, B. K., & Pal, S. 2012).

b) Association rule mining

In the study, some factors which affects the grades of students of Iran and India
has been observed. In such factors, their respective gender, family background,
education level of their parents and their lifestyle has been encountered by

17
asking questions from them (Oskouei, R. J., & Askari, M, 2014). To evaluate
the performance of students and improve the management of educational
institutions, researchers have introduced pre university characteristics of
students which are known as student’s profile and place of secondary school,
final secondary education score, total admission score, and score achieved that
exams. In this research data of University of National and World Economy
(Bulgaria) has been collected and data mining techniques has been applied. In
such data mining techniques naïve bayes and bayes net, nearest neighbor
algorithm, and two rule learners One R and JRIP has been applied on the
dataset (Kabakchieva, D. 2013).

c) Pattern mining
As discussed earlier, it is now clear knowledge is gathered about online
teaching and learning system. Universities through internet are now big source
of student-teacher interaction and source of learning and training. This
technology has made place in the field of educational data mining in which
knowledge diffusion has become research icon for the researchers who belong
to field of data mining, web mining and graph algorithms. The data of online
teaching system is serving best to the researchers who keen to extract
knowledge from big datasets. Researchers are analyzing patterns of online
learning behavior of students and to draw outcomes from these sets of data,
they have been using machine learning algorithms and various data mining
approaches. In a study, researchers used 19,934 servers to identify the behavior
of students in Taiwan who had registered online courses and with the help of
this data, they drawn conclusion after predicting their performance too (Hung,
J. L., & Zhang, K. 2008). In addition, data mining techniques were proven to be
helpful for the course developers, online trainers, and instructional designers. In
their study, they used WEKA and KNIME tools to perform analysis of
descriptive and artificial intelligence. Moreover, for data visualization and
statistical analysis, they used

SPSS. In another study, data mining techniques such as classification,

clustering, relationship mining, prediction and social area networking have
been applied in educational data. (Sachin, R. B., & Vijay, M. S. 2012). Well all

18
of the mentioned approaches are ultimately works to represent the results after
applying over the data. The concept of data visualization has been derived from
visual reasoning (Inoue, S et al, 2017).

d) Text mining
In the past decade, there are number of tools such as WEKA, RapidMiner, R,
KEEL, and SNAPP that have been used to extract text (Useful information)
from the datasets in the field of educational data mining (Baker, R. S., &
Inventado, P. S. 2014). In other research, data mining is also known as
knowledge discovery from databases (Anand, S. S. et al, 1996). In this process,
database techniques are bind with mathematical and artificial intelligence
techniques.

2.4 Commonly used attributes for student performance prediction.

While review of papers from the literature, many data mining approaches have
been found that are applied in academic datasets of different educational
institutions for various purposes. Such approaches have been applied on the
attributes that are considered after analyzing the factors. Those factors are
briefly discussed in following passage.

I. Demographic attributes

From the literature, the importance of demographic attributes upon student’s

performance is inaugurated. According to the researchers, they condemn that
living standard of any student ultimately affect their studies and variation their
performance. Demographic factors include Age, gender, Financial status,
Balance due, Permanent address, Residential address, Guardian (Brother, uncle,
Self), Father qualification, Father Occupation, Full/ Part time student.
Demographic factors equally supports the researchers whether to find the
student attrition or student’s retention. Student’s retention is an obvious success
of educational institutions. In order to sustain from student retention and keep
the performance of students up to the mark, educational institutions better need
to recruit the limited students as per its availability. In a study, the student’s

19
general weighted average, School’s Radial Distance, and school ownership are
taken (Abaya, S. A., 2013) (Gerardo, B. D., 2013)

II. Pre-university attributes

Another factor which has been found during literature survey is pre-university
attributes. These attributes have quite great impact on student’s performance in
any educational institution. Pre-university attributes will be fruitful with respect
to accommodate the student’s interest to map in course recommendation
system. Pre-university attributes include Secondary School Grade, Higher
Secondary Grade, SAT Score, Pre-college, Pre-board, pre-program.
Researchers have used the historical data of students as well to find the attrition
rate of students from academia through cross validation process under the
classification and naïve based methods Guarín, C. E. L., 2015)( Guzmán, E.
L.,2015)( González, F. A. 2015).

III. Institutional attributes

Last factor that has been extracted from literature to improve the performance
of students is institutional attributes. These attributes will support in modeling
the teacher recommender system and future course prediction model in our
study. In institutional factors, Term 1 GPA, Term 2 GPA, Term 3 GPA, and
Term 4 GPA, CGPA, Total Credit hours taken, Total courses taken and Initial
major, current major, current enrollment status are included.

Proposed attributes of students affects their performance in various ways. It not

only helps in identifying the students with low academic performance but also
paves ways to improve their grades and enables recommender systems of
universities to predict the grades of students. With the help of this study,
student carrier can be effected by means of building up and students can easily
pave their way towards their desired destination. This study will also help to
construct the course recommender system which will enable students to select
their courses that are recommended by the proposed system. Another advantage
of this study is highly beneficial economically for parents. With the help of
proposed system, students may not waste their money in selection of courses to

20
repeat. Moreover, student’s academic record can be improved and efficiency of
department management can be raised.

2.5 Literature review summary

Table 2.1: Literature Review Summary

Used
Attributes

Reference Relevant Techniques Tools

Pre-university
Demographic

Institutional
attributes

attributes
10 fold cross
(Guarín, C. E.
Decision Trees, validation
L., Guzmán, E.
Bayesian model, Cost √
L.,González, F.
Classification sensitive
A. 2015)
model
(Abaya, S. Irecruit
Classification through
A.& Gerardo, Application √
C4.5
B. D., 2013) in UNIX
Clustering, Feature
CHAID
(Kovacic, Z. selection thorough
Model, Gian √
2010). cross validation, and
chart
CART classification
CRISP-DM, Neural
networking, Nearest
Kabakchieva,
neighbor classifier, WEKA √
D. (2013).
Rule learner, Decision
tree classifier
(Al-Barrak, M.
E-learning
A., & Al- Data Visualization,
Web Miner, √
Razgan, M. Classification
WEKA
2016).
(Baradwaj, B.
Classification and
K., & Pal, S. Manual √
Decision trees
2012)
Recommend
Matrix factorization
er system
(Elbadrawy et based methods,
based √
al, 2016) regression based
personal
methods
analytics
(Bydžovská, EDM, SNA and WEKA and
√
H. 2013). Collaborative filtering R
(Kaur, G., &
Naïve based and J48
Singh, W. WEKA √
Decision trees
2016)
(Kokina, J.,
Pachamanova, Predictive Modeling, Excel and
√
D., & Corbett, Data Visualization Tableau
A., 2017).

21
Chapter 3

RESEARCH METHODOLOGY
This chapter presents the methodology adopted to address the research
questions presented in chapter 1. The focus of the research is accurate grade
prediction of critical courses of first semester of BS CS students of CUST.
Prediction will help to take appropriate measures to control the attrition rate
and hence will be beneficial for students, parents and university. Targeting the
same objective, appropriate faculty members are also being recommended
based on the results of previous semester. The factors under consideration for
the proposed research are classified into three categories; Demographic, Pre-
university and Institutional. Such researches in the field of educational data
mining have been conducted in foreign earlier. Experiments have been
conducted in local context, Pakistan. For this purpose, all the data have been
collected from Capital University of Science and Technology, Islamabad
Pakistan. After data collection and its pre-processing, some data mining
techniques are being selected like SVM, Linear regression and Non-linear
regression in WEKA.

This study will not only be useful in improving the overall performance of
students but also reduces the attrition rate. Now it is easy to do such things on
the basis of early prediction. Field of data mining has been emerged with the
linkage of natural language processing, artificial intelligence, visual data
analytics, and social data analysis etc (Romero, C., & Ventura, S. 2013)(Zhang,
Y.et al, 2010). This early prediction is computed by using term marks and mid-
term marks of the students along with their pre-university; Intermediate/O-
levels marks and Matriculation marks and demographic factors; gender and
city. This methodology will be discussed briefly in the current chapter. The
related work of this research task has been discussed in detail in chapter 2.
Chapter 3 contains detailed discussion of proposed methodology along with the
data flow diagram that has been given as 3.1

22
 Figure 3.1: Data Flow Diagram of Proposed Methodology

Proposed data flow diagram is constructed on the basis of following set of

symbols.

Table 3.1: Description about Data Flow Diagram

This cylindrical shape represents the raw data that has

been collected for the research and need to be pre-
processed.
This filled arrow is used to represent the flow of tasks
that has been carried out for the experiments.
This symbol represents the working of pre-processing
steps in our data flow diagram.
This oval shaped rectangle represents the algorithms
and techniques that are supposed to apply on the data
set for the experiments.
This shape represents the final results that include the
result prediction and evaluation of results.

23
To conduct the experiments, first of all the data is collected from the CUST
(Capital University of Science and Technology), Islamabad Pakistan. The data
set contains the data of six semesters that belongs to the BSCS program. On the
basis adopted factors, the GPA of 1st semester as well as CGPA of 2nd semester
has been predicted. In addition, performance of teachers against their respective
courses has been measured through ANOVA analysis. The experiments have
been conducted through WEKA. Each research question has been answered
with the help of proposed methodology and experiments.

3.1 Data collection and Pre-processing

As discussed earlier, domain of computer science is being selected for the
experiments and CUST has provided the facility to collect the data from its
sources. The academia of CUST is semester based and there are two semesters
per year. One semester corresponds to the spring semester which starts from
February and ends in June. And the fall intake starts from mid of September
and ends in the end of January. The data initially contained 5 courses of first
semester, such as English-1, Physics, Calculus-1, Pakistan Studies & Islamiat
and ITC/CP.

Table 3.2: Dataset Summary

Term Registered Students

Spring 2014 74

Fall 2014 155

Spring 2015 115

Fall 2015 102

Spring 2016 62

Fall 2016 101

Spring 2017 58

Total 667

24
However, the information has been extracted from a research (Junaid. 2017),
that most crucial courses with respect to attrition or students’ performance are
ITP/CP and Cal-I. So results of only these two semesters are considered and
excluded the rest three.

Collected data is related to the students of first semesters and seven semesters
starting from Spring 2014 (Term 141) till Spring 2017 (Term 171), and also of
second semester students of six semesters starting from Fall 2014 (Term 143)
to Spring 2017 (Term 173). The courses of first semester, as mentioned earlier,
are Cal-I and ITC/CP, and the second semester courses are Cal-II and CP/OOP
from each spring new intakes that belongs to the semester no 1. The initiative
of this research is to identify the factors which contribute to keep the students
retained till the end of the semesters as well those factors that affect the
performance of the students. Following table 3.2 shows the information of the
students comprehensively.

Table 3.2: Summary of Session-wise Attrition

2nd Sem

3rd Sem

4th Sem

6th Sem

7th Sem

8th Sem
1st Sem

5h Sem
Term

141 64% 20% 8% 8% 0% 0% 0% 0%

143 29% 35% 25% 12% 0% 0% 0% 0%

151 60% 30% 9% 0% 0% 0% 0% 0%

153 71% 25% 4% 0% 0% 0% 0% 0%

161 94% 6% 0% 0% 0% 0% 0% 0%

163 100% 0% 0% 0% 0% 0% 0% 0%

Attrition % 70% 19% 8% 3% 0% 0% 0% 0%

25
The dataset contains the institutional factors; Registration Number, Name,
GPA, Internal marks like midterm, and term works marks etc and overall
information of the students such as Student’s registration number, GPA,
Registered courses, Matriculation and Intermediate results. To conduct the
experiments, we have used demographic and pre-university factors of the
students apart from institutional factors. In the data set, the data of term 141,
143, 151, 153, 161, 163, and term 171 are included. The terms of the spring
semesters are associated with the registration numbers of the students such as
the term 141 means spring 2014. Similarly, the data of spring 2015 and spring
2016 are collected and on the basis of three years record of the students. The
file of dataset is initially imported into WEKA after converted into CSV and
ARFF format for the pre-processing and further experimentations have been
conducted respectively.

Missing Data Handling

The problem of missing data generally arises due to absence of data in an

observation for any variable during experiments. Missing data problem also
arises when no information is provided or unavailable for the variables. The
reason of missing data may involve the problem in collection of data by the
researchers or the sources have not provided it completely3.
Missing data has been handled by two ways. Firstly, null values of courses
have been found from other semesters and may be read in late semester in
which students have passed that course. In case of consistent absence of values,
the grades of such course have been supposed by taking the average result of
other courses in respective semester. The missing data can be handled through
different filters in WEKA.

3
https://measuringu.com/handle-missing-data/

26
Noisy Data Handling
While data collection, irrelevant data is called as noise. As the data of BSCS
students of spring 2014 from spring 2016 is collected, therefore during the
collection, data belongs to Software engineering and bioinformatics was
present and then eliminated through filters in WEKA. Noisy data occurred in
the form of errors such as; GPA = “−0.4” and Intermediate or Matriculation
marks = “-455, 544.8” which were handled in Pre-processing of data.

Attribute Selection
After missing data handling and outlier detection phase, final attributes have
been selected on which overall experiments and results depend. Such attributes
have been narrated in table 3.1 that belongs to all three types of factors;
demographic, pre-university, and institutional. One more factor has been
considered to answer our research question that is teacher performance. After
completion of data pre-processing, classification algorithms have been applied.
Attributes are then finalized for the experiments that have been evaluated to
answer the research questions. In the current section, we have discussed
adopted classifiers in detail.

3.2 Classification
Before applying algorithms, factors have been grouped that are earlier used by
foreign researchers. These factors have been compared against the factors used
by the researchers in local context (Pakistan). Moreover, we have expanded our
research work by making the variations of different combinations attributes and
then compared. The comprehensive detail about these factors has been
mentioned in table 3.4.

27
 Table 3.3: Selection of Attributes
Sr.No. Factors Attribute Name
1. Gender
2. Age
3. Residence
4. DEMOGRAPHIC Location
5. Race
6. Father’s qualification
7. Father’s occupation
8 Secondary school grade
9 Higher secondary grade
10 PRE-COLLEGE Pre-college
11 Pre-program
12 SAT score
13 GPA (Term1,
Term2,Term3, Term4)
14 CGPA
15 Financial status
16 Total credit hours taken
INSTITUTIONAL
17 Total courses taken
18 Initial Major
19 Current Major
20 Current enrollment status
21 Teacher methodology
22 Instructor grading
23 Instructor feedback
24 TEAHCER Instructor teaching
methodology
25 Teacher name

28
3.3 Algorithms and Techniques
To evaluate our research by using different classification algorithms; Naïve
Based and J48 has been compared. These classifiers are chosen on the basis of
different reasons i-e; these classifiers supports Categorical and Nominal. For
this purpose WEKA 3.8 tool has been used. The tool is open source available
on the web and is generally used for machine learning algorithms,
classification, training and testing of data. The data is then can be visualize in
the form of graphs as well. In present research, the formed combinations of
attributes have been evaluated to answer our research questions. For this
purpose the data has been converted into csv format to import in the WEKA
and the results from various filters and classifiers have been accumulated.

a) Naïve Based
Naïve based algorithm is comparatively fast algorithm in terms of
classification. It works faster on huge datasets by using Bayes algorithm of
probability. Bayes algorithm generally used to predict the class of unknown
dataset4. Naïve based algorithm works on assumptions to label an item whose
features are known but name is unknown. For example; a fruit is labeled as an
apple if it is round and red in color and its size is 3 inches in diameter. These
features of apple will raise the probability of this fruit that it is an apple.

b) J48
J48 decision tree is used to predict the target variable of new dataset. If dataset
contains predictors or independent variables and set of target or dependent
variables, then this algorithm is applied to extract the target variable of new
dataset5.

c) Linear Regression

Generally linear regression classification algorithm is an approach to identify

the relationship between dependent and independent variable. It is generally
used for predictive analysis and has two main points. One is to check whether
predictor variable does a good job in predicting the expected outcome variable.

4
https://www.analyticsvidhya.com/blog/2015/09/naive-bayes-explained/
5
http://data-mining.business-intelligence.uoc.edu/home/j48-decision-tree

29
Second main thing that linear regression does is the identification of variable
that are significant predictors of dependent variables. At the end, the regression
equation is used which helps to determine the set of predictor which are used to
predict the outcome. In this research, algorithms are being used to compare the
trend and pattern of the factors with other approaches like non-linear regression
and SMO.

3.4 Evaluation of Research Questions

For our research, we have proposed three research questions that have been
evaluated in multiple dimensions. Such dimensions have been explored in
following ways respectively.

RQ1: Whether exiting identified factors for grade prediction are valid in
our local context (Pakistan)?
Our research question number 1 determines the validity of factors in local
context which have been carried out in foreign context by the researchers. The
main reason that distinguishes this research in local context from the foreign
context is that the norms, traditions and culture of every country that varies.
Therefore it is needed to be evaluated that whether the existing factors that
work in the colleges and universities of Canada, USA, England, and Nigeria
etc, are applicable in Pakistan or not. After pre-processing of whole data,
classification algorithms have been applied to analyze the behavior of
demographic, pre-university and institutional factors. According to the results,
it has come to known that these factors are also valid in Pakistan because of
nearest difference in accuracies. Detailed results have been discussed in chapter
no 4.

For our research, data being collected from term 141 to term 171 which consist
of 667 students. Then three types of exiting attributes to predict the
performance of students have been taken. First type of attribute is
demographic; Gender, Age, Residence, Location, Race, Father’s Qualification
and have considered only Gender and Age of the student. Second type of
attribute is Institutional; GPA(Term1, Term2, Term3, Term4, CGPA, Financial
status, Total credit hours taken, Total courses taken, Initial Major, Current

30
Major, Current enrollment status, Teacher methodology and Midterm was
considered, Term Marks for the prediction of final CGPA.

Third type of attribute is Pre-qualification; Secondary school grade, higher

secondary grade, Pre-college, Pre-program, SAT score and Marks of FSC and
MATRIC are considered. Among these attributes, specific attributes

During pre-processing, it has been found the total number of students who
could not continue their program and left the course incomplete. After pre-
processing, there were total 660 remained and 7 out of 667 who left the course.
Two types of errors occurred in this dataset. The first was missing data and
second was incomplete data which has been corrected in the phase of pre-
processing.

In Table 3.5, Attributes selection to dataset for experiments has been shown

Table 3.5: Attributes selection to Dataset

Mid Physics
Mid Eng-1

Mid Cal-1
Mid ITC
Gender

Studies
Sr. No.

Matric

GPA
F.Sc
City

1 Male Rawalpindi 51 64 10 10 16 9 16 2.51

2 Male Islamabad 61 56 9 9 8 8 8 2.22

3 Male Rawalpindi 51 64 7 7 13 17 13 3.15

4 Male Islamabad 48 55 9 9 12 7 12 3.26

5 Male Islamabad 77 58 12 12 17 18 17 3.07

6 Male Islamabad 74 60 9 9 16 18 16 2.74

7 Male Rawalpindi 84 70 6 6 12 7 12 2.86

8 Male Islamabad 64 51 7 7 12 6 12 1.9

9 Female Islamabad 61 56 8 8 11 6 11 1.83

31
 RQ2: Which factors help in accurate prediction of student’s GPA of first
semester?
Some specific attributes form all three factors; demographic, institutional and
pre-university have been selected. The combinations of those attributes are
constructed and then filters have been applied respectively. There are total 21
unique combinations of attributes presented in table 3.6.

Table 3.6: Pattern of Combination of Attributes

Sr.
No. ATTRIBUTES
Demographic, Pre-Qualification, Institutional (Result all subject of 1st
1 semester)
Demographic, Pre-Qualification, Institutional (Result Eng-1, Cal-1, ITC,
2 subject of 1st semester)
Demographic, Pre-Qualification, Institutional ( Result Eng-1, Cal-1 subject
3 of 1st semester)
Demographic, Pre-Qualification, Institutional (Result Eng-1 subject of 1st
4 semester
5 Demographic, Pre-Qualification, Institutional (Cal-1 subject of 1st semester)
Demographic, Pre-Qualification, Institutional (Result ITC, subject of 1st
6 semester)
Demographic, Pre-Qualification, Institutional (Result Cal-1, ITC subject of
7 1st semester)
Demographic, Pre-Qualification, Institutional (Result Eng-1, ITC subject of
8 1st semester)
9 Demographic, Pre-Qualification, Institutional(GPA )
10 Demographic, Pre-Qualification
11 Demographic
12 Pre-Qualification
13 Pre-Qualification, Institutional (Result all subject of 1st semester)
14 Demographic, Institutional ( Result all subject of 1st semester)
15 Institutional (Result Eng-1, Cal-1, ITC, subject of 1st semester)
16 Institutional (Result Eng-1 subject of 1st semester)
17 Institutional ( Result Cal-1 subject of 1st semester)
18 Institutional (Result ITC, subject of 1st semester)
19 Institutional (Result Eng-1, Cal-1 subject of 1st semester)
20 Institutional (Result Cal-1, ITC subject of 1st semester
21 Institutional (Result Eng-1, ITC subject of 1st semester)

32
RQ3. Is it possible to improve the performance of a student by
allocating of appropriate teacher for a subject?
There is a fact that is normally faced during the research period is; the
personality of teacher affects the performance of the students. If the
background of the teacher is already known with the help of prediction system,
it might become feasible for the department to recommend the best suitable
teacher to the student up to his level of interest. And this phenomenon can
surely boost the performance of the student. To answer the question number 3,
experiments have been performed using two approaches have been used. And
these approaches Statistical experiments and Predictive experiments.

a) The research has used combination of techniques comprising simple

mean and the prediction based on teacher. The focus of this question is
to find the most suitable teachers for Cal-I and ITP as these two subjects
are considered the first semester with respect to the academic
performance or attrition rate. Following calculations have been
performed for the Cal-I teachers with respect to the prediction
experiments:

1. Average of prequalification is taken

2. Average of actual Cal-I marks based on the Cal-I teacher

3. Average of actual Cal-II marks based on the Cal-I teacher

4. Prediction of the Cal-I and ITC/ITP is taken on the basis of

prequalification

5. Predict Cal-ll marks on the basis Cal-I and average is taken

b) ANOVA data analysis

ANOVA is a statistical technique which is used to measure the difference in a
scale level dependent variable by a nominal variable which comprises of two or
more types of categories. Normally ANOVA test is applied to find out the
significant difference among groups. Similarly we will be applying this test to
find the difference between the performances of teacher’s subject wise in our

33
case of experiments6.
c) Z-score data analysis
Unlike ANOVA (analysis of variance), Z-score is usually applied on three or
more means. A Z-score is a type of hypothesis test which is a way to find
whether the results obtained from a test are valid or need to be repeated. For
example, if someone said they had found a new drug to cure the cancer, one
would want to be sure it was probably true. Similarly, in our research, we have
applied Z-test to compare the performance of teachers particularly with respect
to their subject. Z – Test will exploit the likelihood that the obtained results are
true or not. A Z-test is generally used when the data is approximately normally
distributed in the form of pairs7.

6
http://www.statisticssolutions.com/manova-analysis-anova/
7
http://www.statisticshowto.com/z-test/

34
Chapter 4

RESULTS & EVALUATIONS

Researchers have made their significant contributions to introduce such factors

in the field of educational data mining that are quite beneficial for the
educational institutions to mine hidden patterns from the academic database.
Educational data mining is broadened field in which Online learning
management system as well as physical institution’s datasets are used. One
limitation of this approach is to get the access of the dataset from the
administration in order to perform analysis and conduct research. The dataset
have been collected from Capital University of Science and Technology,
Islamabad and applied existing factors on this dataset to answer our proposed
research questions. Following sections of this chapter will present the overall
picture of obtained outcomes in detail.

Data collection and pre-processing are the initial steps towards the analysis of
the research. For this purpose, the demographic data has been collected from
the registrar office of the CUST which contains student’s gender and city. With
the co-operation of administrative resources of CUST, the data related to pre-
university and institutional factors has been collected. Institutional and pre-
university factors contain the data about mid-term marks, term-marks, GPA,
intermediate marks and matriculation mark. University portal provided
privilege to access this data of the students to gather the sufficient dataset for
our experiments.

The gathered data came into many formats such as PDF, word and excel. For
the pre-processing, firstly this data set brought into the single file; Excel.
Figure 4.1 represents the format of files and transformation of all files into
single one. The collected data set arises in three formats; word, PDF and excel.
After collection of dataset, it became mandatory to transfer the data into single
file from all the files.The purpose of excel file is to import it into the WEKA to
apply filters for the pre-processing. The final versions of dataset in the form of

35
excel is then converted into CSV format and imported in to the WEKA. Such
filters have been applied that are discussed in subsections of section 4.1.

Figure 4.1: Formats of Data

In context of missing data, lacking attribute values, lacking certain attributes of
interest, or containing only aggregate data instead of complete information are
found. Incomplete information may lead the research towards in accurate
outcomes that might affect the overall experiments. Therefore, the missing data
problem has been handled in a way that no attribute remained null or
incomplete.

Table 4.1: Occurrence of Missing Values

Final Exam

Mid Term
Weighted

Teacher
Gender
Sr. No.

Work
Term
Total
ITC

1 Male B- 75.95 76 25.2 15 21.67 D

2 Male D 50.27 50 17.6 7.25 14.8 C

3 Male C 66.93 67 22.8 8.25 21.65 A

4 Male

5 Male B- 76.43 76 21.6 13.75 27.27 D

6 Male D 50.19 50 14.4 10.25 10.95 B

7 Male A 89.93 90 34.8 14.25 30.57 D

36
The methods of data mining behave differently in the way that they treat
missing values. Normally, they ignore the missing values, or exempt those
records which contain missing values or either replace missing values with the
mean, or conclude missing values from existing values. Missing
Values Replacement Policies include the number of strategies such as; ignore
the records with missing values. Following example shows the occurrence of
missing values in the dataset. After the handling of missing data by taking the
average of last three records of students and filled the missing one, complete
dataset is gained. Further, this dataset has been used for experiments to
evaluate all research questions.

Table 4.2: Handling of Missing Values

Term Work
Final Exam

Mid Term
Weighted

Teacher
Gender
Sr. No.

Total
ITC

1 Male B- 75.95 76 25.2 15 21.67 D

2 Male D 50.27 50 17.6 7.25 14.8 C

3 Male C 66.93 67 22.8 8.25 21.65 A

4 Male C+ 72.91 73 19.2 18 27.75 A

5 Male B- 76.43 76 21.6 13.75 27.27 D

6 Male D 50.19 50 14.4 10.25 10.95 B

7 Male A 89.93 90 34.8 14.25 30.57 D

Outliers are an observation point that is found to be distant from other

observation points. An outlier may occur due to the change in the measurement
or it may indicate experimental error; the latter are sometimes excluded from
the data set. After handling noise from the dataset, following result is obtained.

37
Attribute Selection

Appropriate selection of attributes is very important task in the process of pre-

processing. It ultimately affects the performance of results and may lead to
inaccurate results, if not chosen carefully. That’s why we have classified the
attributes into two types; Nominal and Categorical. On these two types of
attributes researcher have selected the attributes that best suit to the nature of
these two types of attributes.

Sampling
Data sampling is a statistical analysis technique used to select, manipulate and
analyze a representative subset of data points in order to identify patterns and
trends in the larger data set being examined. For example: term
141,143,151,153,161,163 dataset has been used as a training data and term 171
is test data because prediction of result of corresponding students is required to
be computed.

4.1 Final Attribute Selection

From the literature, there are plenty of attributes found and some of them have
been stated in the following table. This table belongs to the table presented in
chapter no 3 but used in the current chapter with the addition of column named
as selected attributes. These attributes have been marked to distinguish from
each other. Then final attributes have been acquired for the experiments to
fulfill the requirements of our results.

4.2 Result of Evaluation of Research Questions

With the help of adopted algorithms for the experiments in WEKA, the
proposed research questions have been tried to answer. These answers make
this research methodology validate and complete. Classification and linear
regression has been selected to measure the analysis. The results of
experiments have been discussed in detail in this section for each research
question.

38
RQ1: Whether existing identified factors for grade prediction are valid in
our local context (Pakistan)?
As existing factors along with their attributes have been stated in above
sections, we have drawn the combinations of local factors with the existing
factors. We found a slight variation between both types

of attributes. The comprehensive results of such variation have been arranged

as the answer to each of three questions. The format of attributes has been
presented in table 4.3 which are used in experiments.

Table 4.3: Attributes table

Mid Pak-Studies
Mid Physics
Mid Eng-1

Mid Cal-1
Mid ITC
Gender
Sr. No.

Matric

GPA
City

FSc

1 Male Rawalpindi 51 64 10 10 16 9 16 2.51

2 Male Islamabad 61 56 9 9 8 8 8 2.22

3 Male Rawalpindi 51 64 7 7 13 17 13 3.15

4 Male Islamabad 48 55 9 9 12 7 12 3.26

5 Male Islamabad 77 58 12 12 17 18 17 3.07

6 Male Islamabad 74 60 9 9 16 18 16 2.74

7 Male Rawalpindi 84 70 6 6 12 7 12 2.86

8 Male Islamabad 64 51 7 7 12 6 12 1.9

9 Female Islamabad 61 56 8 8 11 6 11 1.83

As an input we have imported the .CSV into the WEKA that contains
information about each attribute against every student. The input data contains
the data of students of BSCS department from term 141 to term 163. The

39
dataset of CUST (Capital University of Science and Technology) has been used
for this purpose. Among those factors, attributes like Pervious Course
Marks/Grade, GPA, SSG, HSSG, Gender, City has been used. Afterwards, we
have applied well-renown classification algorithms i-e; Naïve Bayes and J48 on
the dataset.

Table 4.4: Classification of attributes and algorithms

Reference Attributes Class Label Naïve Bayes J48

Student
Demographic, Per-
Qualification, 1st
semester 5 Subjects
Grade CGPA Grade 69.12% 72.39%

Student
Data mining Demographic, High
approach for School Background,
predicting student Scholarship, Social
performance (2012) network interaction CGPA Grade 76% 73%

Comparison of Midterm marks all

Classification subjects GPA Grade 65.6% 80.98%
Techniques for
predicting the
performance of
Students Academic Internal assessment,
Extra-Curricular
Environment(2014) activities GPA Grade 66% 73%

Student
Predicting Student Demographic, Per-
Performance: A Qualification, CGPA Grade 49.90% 63.19%
Statistical
Student
and Data Mining Demographic, High
Approach(2013) School Background CGPA Grade 50% 65%

In table 4.4, we have summarized the attributes of each factor with

corresponding to their accuracies. These attributes have been assigned with
class label and the reference of each attribute has been given as well. During
the literature review, we have found the accuracy of each attribute in their
respective scientific research publication and compare the accuracy of our

40
results. The inferences show the slight variations which is quite considerable
with the correspondence to our research. The existing attributes have been
shown in black color in the summary table and attributes in red color are local
attributes.

The summarized results have been shown in the figure 4.2 in which bars of
different colors on X-axis shows the presence of attributes and y-axis shows the
interval of accuracy.

Figure 4.2 Comparison of local and Existing attributes

Our Findings narrate that existing identified factors for grade prediction are
considered to be comparable with the local factors because of low variation in
the result of their accuracies. According to the results, Pervious Grade, Internal
Marks, HSSC were found to be effective for grade prediction.

RQ2. Which factors help in accurate prediction of student’s GPA of first

semester?

Our second research question is the extraction of those factors that helps in
accurate prediction of GPA of students belong to 1st semester. To answer the
research question, we have transformed this question into following two ways.

41
 1. Based on Mid term marks

In this phase of experiment, we have combined the attributes including

marks of mid term to predict the GPA of the first semester. our prediction
leads to the positive focal point where our proposed combinations worked
well.

2. Based on Pervious Grade

In this phase of experiment, we have interpreted our attributes to predict the

CGPA of the students. The CGPA of 2nd semester has been computed on the
basis of previous academic record.

After compilation of attributes, their accuracies have been ranked that is

presented in table 4.5 is descending order.

Information Gain Ranking Filter

Table 4.5: Ranked Attributes

Ranked Attributes
0.2277 Cal-I MID
0.1903 Cal-II MID
0.1778 ITC MID
0.1684 Physics MID
0.1548 Pak-Studies MID
0.1514 FSC
0.1006 City
0.0992 Matric
0.0791 Gender
According to results, picture states the prediction of CAL-I MID and CAL-II
MID remained better in terms of accuracy. Then the accuracies gradually
decreases from ITC MID to GENDER. Based on the corresponding accuracies
of attributes, we may inference the importance of acquired ranking information
gain of all the instances.

42
1- First Semester GPA Prediction

We have expanded our answer for the research question number 2. With the
help of “GPA Prediction for the first semester” we became able to evaluate our
results through WEKA. We covered the list of attributes narrated in table 4.6 in
which the combination of attributes are defined in column 1. There are four
combinations of attributes which belong to the three factors; Pre-Qualification,
Demographic and Institutional. The performance of J48 is considerably better
then Naïve Bayes that has been shown in table as well as presented in figure
4.3.

Table 4.6: GPA Prediction for the first semester

Attributes Class Label Naïve Bayes J48

Pre-Qualification, Institutional (Internal

marks only Midterm) all subject of 1st
semester GPA Grade 64.62% 81.39%

Institutional (Internal marks only

Midterm all subject of 1st semester) GPA Grade 65.64% 80.98%

Demographic, Pre-Qualification,
Institutional (Internal marks only
Midterm all subject of 1st semester) GPA Grade 66.46% 79.55%

Demographic, Pre-Qualification,
Institutional (Internal marks only
Midterm Eng-1, Cal-1 ,ITC subject of
1st semester) GPA Grade 69.33% 78.12%

According to the results, the performance of Demographic, Pre-Qualification,

Institutional (Internal marks only Midterm Eng-1, Cal-1, ITC subject of 1st
semester) remained better as compared to the other three combinations of
factors when Naïve Bayes classifier applied. On applying J48 classifier, the
performance of Pre-Qualification, Institutional (Internal marks only Midterm)
all subjects of 1st semester remained better comparatively to predict the grades
of first semester.

43
 Figure 4.3: GPA Prediction for the 1st semester
2- CGPA of second Semester

To predict the CGPA of 2nd semester, we have defined the following narrated
set of experiments as summarized in table 4.7. We have merged all the grades
of previous semester along with the adopted factors. These combinations and
their accuracies on the basis of Naïve Bayes and J48 have been presented in
table 4.4.

Table 4.7: CGPA Prediction for the second semester

Class Naïve
Attributes Label Bayes J48

Demographic, Pre-Qualification,
Institutional(Subject Grade Cal-1 subject of CGPA
1st semester) Grade 67.08% 74.44%

Demographic, Pre-Qualification,
Institutional(Subject Grade ITC subject of 1st CGPA
semester) Grade 68.10% 74.44%

Demographic, Pre-Qualification, Institutional

(Subject Grade Cal-1, ITC subject of 1st CGPA
semester) Grade 70.55% 73.82%

Demographic, Pre-Qualification, Institutional

(Subject Grade Eng-1, Cal-1 subject of 1st CGPA
semester) Grade 65.44% 72.60%

44
These tabular results have been presented as figure 4.4. Like 1st semester GPA
prediction, we have used same classifiers for the CGPA prediction for 2md
semester. The only change lies in the combinations of attributes. According to
the accuracy of each combination, we reached to the point that, the accuracy of
Demographic, Pre-Qualification, Institutional (Subject Grade Cal-1, ITC
subject of 1st semester) was high whereas when J48 classifier was applied, the
prediction of Demographic, Pre-Qualification, Institutional(Subject Grade Cal-
1 subject of 1st semester) and Demographic, Pre-Qualification,
Institutional(Subject Grade ITC subject of 1st semester) remained better than
other two combinations with the minor difference i-e; Demographic, Pre-
Qualification, Institutional (Subject Grade Cal-1, ITC subject of 1st semester)
and Demographic, Pre-Qualification, Institutional (Subject Grade Eng-1, Cal-1
subject of 1st semester).

Figure 4.4: CGPA Prediction for the second semester

Our Findings enlighten the following assumptions.

 For Grade Predictions, the Attribute Evaluator has computed the

Information Gain and ranked the accuracy of each attribute accordingly.

45
 Then it is found that the subjects of First semester i-e; Cal-1, Eng-1,
ITC have high value of information gain and played an effective role in
order to predict the accurate grades of students.

 As we came to know from the situation of dataset that the Attrition rate
is high in first semester. The purpose of the research is to reduce the
attrition rate, so we have considered these subjects to predict the
attrition rate of students as well as student’s performance.

 Results states that Demographic attributes does not play significant role
in grade prediction of students.

 In addition, internal marks like Mid Terms are used for early grade
prediction that has high accuracy and also helps to minimize the
attrition rate.

RQ3. Is it possible to improve the performance of a student by allocating of

appropriate teacher for a subject?
The RQ3 has been addressed with a combination of techniques comprising
simple mean and the prediction based on teacher. As it has been discussed in
the precious chapter that the focus of this question is to find the most suitable
teachers for Cal-I and ITP as these two subjects are considered to be critical for
the first semester with respect to the academic performance or attrition rate. To
evaluate the research question no 3, we have adopted the following set of
experiments. Initially, the dataset has been imported into WEKA that contains
the information of students from term 141 to term 163. The set of attributes
consist of the Pervious Core and Elective Course Internal Marks, GPA, SSG,
HSSG, Gender, City. After finalizing the set of attributes, we have applied
following proposed techniques.

• Predictive Approach: Classification algorithm: Linear Regression

• Statistical Approach: Simple average , Z-Score analysis , ANOVA

data analysis

The experiments have been conducted subject wise for the individual course
grade prediction. On the basis of course grades, accumulated teacher

46
performance has been computed. Furthermore, following calculations have
been performed for the Cal-I teachers:

1. Average of prequalification is taken

2. Average of actual Cal-I marks based on the Cal-I teacher

3. Average of actual Cal-II marks based on the Cal-I teacher

4. Prediction of the Cal-I is taken on the basis of prequalification

5. Predict Cal-ll marks on the basis Cal-I and average is taken

Pre-Qualification vs Maths Courses

The comparison of Maths course and Pre-qualification has been made in this
phase of experiment. The teachers have been labeled alphabetically and the
overall average of the students has been given against every teacher in table
4.8.

Table 4.8: Pre-Qualification VS Maths Courses

Teacher Pre-Qualification average Cal-I average Cal-II average

A 58.37142857 58.7429 64.4269

B 61.70231214 64.8324 68.03595

C 59.92207792 57.91 66.5098

D 60.17073171 56.1951 64.18938

E 58.96428571 73 72.75

F 58.95588235 57.2059 54.42167

G 60.73512748 61.8491 63.43205

H 59.82369942 61.3628 68.46664

The result shows that the performance of teacher “B, D and E” is found to be
better in Pre-qualification combination. With respect to the Cal –I performance
of teacher “B, G and H remained better. On using the Cal-II attribute, the
performance of teacher “B, C, and D” found to be better than other teachers.
These results have been shown graphically in figure 4.5.

47
 Figure 4.5: Pre-Qualification vs Maths Courses
Average of Cal-ll Teachers
In this case, we have predicted the the performance of teahcers who teach Cal-
II and compared with the actual performance of the teachers in Cal-II.
According to the results, we have found significantly unexpected outcomes.
There is a minor variation between actual Cal-II and predicted Cal-II. The
results have been narrated in table 4.9 and presented in figure 4.6.

Table 4.9: Average of Cal-ll teachers

Teacher Cal-II ave Pred-Cal-II ave

A 64.42689655 62.16017221
B 68.03594595 69.39823039
C 66.50980392 63.70893212
D 64.189375 64.42028569
E 72.75 71.255395
F 54.42166667 56.0806705
G 63.43205128 67.83774562
H 68.46663551 67.88178679

48
 Figure 4.6: Average of Cal-ll teachers
Average of Cal-I teachers

The average of Cal-I teachers has been compared with the results of actual and
predicted values. There is another attribute “Pre-qualification” that has been
used in this case. The obtained results have been given in the table 4.10 as well
as presented in figure 4.7.

Table 4.10: Average of Cal-I teachers

Teacher Pre-Q comb Pred-C1-Preq Cal-I ave

A 58.37142857 60.6696821 58.7429
B 61.70231214 61.50918169 64.8324
C 59.92207792 61.91539229 57.91
D 60.17073171 59.51882925 56.1951
E 58.96428571 60.379367 73
F 58.95588235 59.785871 57.2059
G 60.73512748 63.12452977 61.8491
H 59.82369942 61.70135392 61.3628

49
 Figure 4.7: Average of Cal-I teachers

Average of Maths Teachers

Like cal-I and Cal-II the overall performance of math’s teachers have been
computed against each teacher who teaches Cal-I and Cal-II. The information
in table 4.11 narrates that the performance of teacher “B and E” remained better
than other teachers who teach same subjects. This information has been
presented in figure 4.8.
Table 4.11: Overall Average of Maths Teachers

Teacher Over All Performance

A 60.87421589
B 65.09561403
C 61.99324125
D 60.89886433
E 67.26980954
F 57.2899981
G 63.39571083
H 63.84725513

50
 Figure 4.8: Overall Average of Maths Teachers

After the computation of results of Cal-1 and Cal-II, we became able to find the
teacher “E” whose performance found to be better and able to recommend
teaching Cal-I and Cal-II. Similarly, we will compute the results of other
subjects which are to be taught in the 1st semester and then computed similarly.
Firstly, researchers have computed the performance of every teacher of each
subject which is being taught in first and second semester and gained the
results. The results have been interpreted for every subject with respect to its
related teacher. These results have been shown significant average of all
students in Cal-1 & Cal-2 with Cal-1 teacher name. After the computation of
results for Cal-1 and Cal-II, it became clear to find that the teacher “E” whose
performance found to be better and able to recommend teaching Cal-I and Cal-
II then teacher B have better performances and further teacher B and H
performs better among all other teachers.

Further the average of these teachers will be checked whether it is significant or

not significant by applying statistical test “ANOVA”.

ANOVA data analysis on Cal-I

ANOVA analysis is used for group data we have many teacher for Cal-1 &
Cal-2 subjects. The prediction shows that teacher average performance is

51
significant. To prove these results, we applied ANOVA analysis on this dataset.
We have applied ANOVA (Analysis of Variance) test on teacher’s average. By this
way, results indicate the performance of all Cal-I teacher on the basis of group.

Table 4.12: ANOVA Analysis on Cal-1

Anova: Single Factor

SUMMARY

Groups Count Sum Average Variance

Row 1 5 303.5967 60.71935 15.95621

Row 2 5 330.7894 66.15789 3.27359

Row 3 5 313.1846 62.63691 19.02353

Row 4 5 292.9438 58.58876 27.08998

Row 5 5 357.9515 71.5903 2.338546

Row 6 5 274.268 54.85359 20.50589

Row 7 5 313.5665 62.7133 1.073095

Row 8 5 318.981 63.79621 18.29478

ANOVA

Source of
SS Df MS F P-value F crit
Variation

Between
872.7731 7 124.6819 9.273853 3.14E-06 2.312741
Groups

Within Groups 430.2225 32 13.44445

Total 1302.996 39

In this case F > F crit has been shown as 9.273853 >2.312741. Researchers
reject the null hypothesis and concluded that there are significant differences
between the methods because all 11 methods don’t have the same mean.

52
Z-Score Data Analysis
Now Z-score test is applied to evaluate the performance of every pair of
combination of teacher’s subject wise. Generally, Z-test is applied on two
populations to compare its proportion. In this case, we are supposed to compare
the performance of teachers of the same course which is to recommend to the
student. Researchers have classified each Z-test with respect to the subject and
then have considered the results of Z-score with respect to Cal-1. All the
teachers who teach Cal-I have been indicated with alphabetic characters i-e; A,
B, C, D, E, and F. Pairs such as teacher A with other teachers of Cal-1 are
made for experiment. Moreover, for the comparison, teacher’s data has been
classified as same number of students and different number of students.

Table 4.13: Two Samples for Mean with respect to teacher “E”

z-Test: Two Sample for Means

Variable Variable
1 2
Mean 74.94286 58.74286
Known Variance 102.1429 333.391
Observations 15 35
Hypothesized Mean
Difference 0
Z 4.008258
P(Z<=z) one-tail 3.06E-05
z Critical one-tail 1.644854
P(Z<=z) two-tail 6.12E-05
z Critical two-tail 1.959964

Z-score data Analysis applied on the Cal-1 teacher “E” with Teacher “B” on
the basis different number of student. Value of Z-critical to tail < Z concluded
that the performance of teacher “E” is not significant.
Then teacher “B” has been paired with teacher “A” on the basis same number
of students. The value of Z-critical > Z concluded that significant performance

53
of teacher “A”. The results have been shown in table 19. Z-test has been
applied in both cases and results have been compared either the performance of
teachers remain same or vary. By this way, results indicate the performance of
every teacher of Cal-I on the basis of combinations. The result of z-test will
evolve around this statement “If z< z critical two tail then performance
difference is not significant”. If z > z critical two tail then performance
difference is significant.

Table 4.14: Two samples for means with respect to teacher “B”

z-Test: Two Sample for Means

Variable Variable
1 2
Mean 74.94286 58.74286
Known Variance 102.1429 333.391
Observations 15 35
Hypothesized Mean
Difference 0
Z 4.008258
P(Z<=z) one-tail 3.06E-05
z Critical one-tail 1.644854
P(Z<=z) two-tail 6.12E-05
z Critical two-tail 1.959964

Here, researchers have applied Z-score data Analysis on the Cal-1 teacher “B”
with Teacher “A” on the basis different number of students who enrolled in
different terms from term 141 to term 171. In this case, value of Z-critical two
tails is greater than the value of Z which shows that performance of teacher “B”
is not significant.

Then teacher “B” has been paired with teacher “A” on the basis same number
of students and Z-test has been applied. The results shows that value of z-
critical shows the significant performance of teacher “A” as it is greater than
the value of Z.

54
Teacher B with different and same number of Students

Table 4.15: Teacher B with same number of Students

Cal-1 Teacher Z z Critical two-tail

A 4.008258 1.959964
C 3.370917 1.959964
D 5.288815 1.959964
F 4.463791 1.959964
G 4.119857 1.959964
H 2.997759 1.959964
E 3.630363 1.959964

Z-score data Analysis applied on the Cal-1 teacher “B” with all Teachers on the
basis different number of student. The value of Z-critical tow tail > Z conclude
that performance of teacher “E” is not significant with Teacher “A”, ,”C”, “D”,
“F”, “G”, “H”.,”E”
Table 4.15: Teacher B with different number of Students

Cal-1 Teacher Z z Critical two-tail

A 3.687091 1.959964
C 1.235961 1.959964
D 3.480026 1.959964
F 2.829936 1.959964
G 3.218075 1.959964
H 1.61421 1.959964
E 1.271677 1.959964

Z-score data Analysis applied on the Cal-1 teacher “B” with all Teachers on the
basis same number of student. The value of Z-critical tow tail > Z conclude
that performance of teacher “B” is not significant with Teacher “A”, “B”,”C”,
“D”, “F”, and significant with Teacher “G”, “H”. The ANOVA test has been
applied on the cal-I only. The results have been computed for the rest of the
courses similarly.

55
Pre-Qualification vs Programming Courses
In table 4.17, we have narrated the results of all programming courses which
are taught in 1st and 2nd semester. These averages of courses have been
compared with the Pre-Qualification.

Table 4.17: Comparison of all programming courses

ITC /ITP OOP/CP Simple

Teacher Pre-Q comb Simple Average Average
A 57.170543 61.48148148 65.27777778
B 61.4620758 53.7037037 62.80851064
C 60.99346693 65.21428571 69.5
D 61.04028149 61.81578947 72.39215686
E 60.28896102 71.40625 67.91666667
F 59.72643733 77 68.58333333
G 59.72643733 66.1871345 66.73484848
H 61.887987 66.9375 59.07142857
I 59.69578266 64.16129032 68.7826087
J 59.48424926 62.55882353 70.68421053
K 58.0787338 67.575 63.72222222
L 60.18534971 67.44736842 68.3125

Figure 4.9: Comparison of ITC/ITP teacher

According to the figure 4.8, the performance of all of the teachers has been

56
computed. The performance of teachers with respect to courses average
changes.

Average of ITC/ITP & CP/OOP Student

The average of students who studies ITC/ITP has been computed against every
teacher who teaches these subjects. Then these results have been shown in table
4.18 and presented in figure 4.10.

Table 4.18: Comparison of Actual and Predicted Grades of ITC/ITP

Pre-Qualification Predict ITC ITC /ITP

Teacher average /ITP average Simple average

A 57.170543 60.90678396 61.48148148

B 61.4620758 63.21581404 53.7037037

C 60.99346693 61.1117056 65.21428571

D 61.04028149 68.16813274 61.81578947

E 60.28896102 70.68016263 71.40625

F 59.72643733 69.13777079 77

G 59.72643733 63.65621403 66.1871345

H 61.887987 72.60489569 66.9375

I 59.69578266 65.2340331 64.16129032

J 59.48424926 70.26083656 62.55882353

K 58.0787338 61.80150053 67.575

L 60.18534971 65.17485658 67.44736842

57
 Figure 4.10: Comparison of Actual and Predicted Grades of ITC/ITP
Then the actual and predicted OOP/CP grades have been predicted against
every teacher. The results shows that there is a slight variation in the results
that has been stated in table 4.19 and presented in figure 4.11

Table 4.19: Comparison of Actual and Predicted Performance of OOP/CP

Teachers

Teacher OOP/CP ave Pred- OOP ave

A 65.27777778 65.99605885
B 62.80851064 58.81069353
C 69.5 67.8661911
D 72.39215686 66.23364694
E 67.91666667 72.16535104
F 68.58333333 73.23220667
G 66.73484848 67.4231563
H 59.07142857 65.26328164
I 68.7826087 67.14102048
J 70.68421053 67.37969774
K 63.72222222 66.97798275
L 68.3125 66.77153434

58
 Figure 4.11: Comparison of Actual and Predicted Performance of
OOP/CP Teachers
The overall performance of OOP teachers have been presented in table 4.20 as
well in figure 4.12.

Table 4.20: Overall Performance of OOP Teachers

Teacher Over All Performance

A 62.16653
B 60.00016
C 64.93713
D 65.93
E 68.49148
F 69.53595
G 64.74556
H 65.15302
I 65.00295
J 66.07356
K 63.63109
L 65.57832

The results show that the performance of teacher “E & F” is considerably high
and better than other teachers. Therefore, based on the acquired results, we may
recommend these two teachers to teach the courses in respective semesters.

59
 Figure 4.12: Overall performance of OOP Teachers

These results have been shown significant average of all students in ITC/ITP&
CP/OOP with ITC/ITP teacher name. During the computation of results of
ITC/ITP and CP/OOP, we have found teacher “F” performance better and able
to recommend teaching ITC/ITP and CP/OOP.

ANOVA data analysis on ITC/ITP

ANOVs analysis is used for group data we have many teacher for Cal-1 & Cal-
2 subjects. The average prediction shows that teacher average performance is
significant. To prove this result applies ANOVA analysis on this dataset.

If F > F crit, we reject the null hypothesis. This is the case 8.542703
>1.952212. Therefore, we reject the null hypothesis. The means of the 11
populations are not all equal. At least one of the means is different. Hence we
conclude that there are significant differences between the methods (i.e. all 11
methods don’t have the same mean).However, the ANOVA does not tell you
where the difference lies. Therefore it is preferred T-Test, Z-Score test each
pair of means.

60
 Table 4.21: ANOVA Data Analysis on ITC/ITP

Anova: Single Factor

SUMMARY

Groups Count Sum Average Variance

Row 1 6 370.458 61.743 9.09211

Row 2 6 346.3743 57.72906 16.59501

Row 3 6 404.7156 67.4526 4.601526

Row 4 6 397.287 66.21451 25.45386

Row 5 6 421.2417 70.20696 3.492003

Row 6 6 441.7835 73.63059 19.18218

Row 7 6 400.4038 66.73397 0.377307

Row 8 6 382.6283 63.77138 24.15625

Row 9 6 394.865 65.81084 6.082654

Row 10 6 392.0845 65.34741 12.9729

Row 11 6 382.7968 63.79946 15.06479

Row 12 6 407.9563 67.99271 0.222976

ANOVA

Source of
Variation SS df MS F P-value F crit

Between
Groups 1075.12 11 97.73818 8.542703 7.87E-09 1.952212

Within
Groups 686.4678 60 11.44113

Total 1761.588 71

Z-Score data analysis on ITC/ITP

As discussed earlier, Z-test is a measure to analyze the performance of two
attributes which have either equal or different number of population. Here, we

61
have applied Z-test on the combinations of teachers who teach ITC/ITP. We
have made combinations similarly as we made for the teachers of Cal-1 and
Cal-II. In table ..., Z-test has been applied on each pair of combination and
tested on two cases. First is to compare the performance of teachers on same
number of students and second is to compare the performance of teachers on
different number of students. Z-score data Analysis ITC/ITP has been applied
on the pair in which performance of teacher “F” with Teacher “A” on the basis
different number of students has been compared. The results show that there is
significant performance of teacher “F” with teacher “A”.

Table 4.22: Z-Score Data Analysis Teacher F with A different number of

Students

z-Test: Two Sample for Means

Variable 1 Variable 2

Mean 77 61.48148

Known Variance 260.2985 278.7682

Observations 67 27

Hypothesized Mean Difference 0

Z 4.116767

P(Z<=z) one-tail 1.92E-05

z Critical one-tail 1.644854

P(Z<=z) two-tail 3.84E-05

z Critical two-tail 1.959964

Z-score data Analysis has been applied on the pair of ITC/ITP teacher “F” with
Teacher “A” on the basis same number of students and results show that there
is a significant performance between teacher “F” and teacher “A”.

62
 Table 4.236: Z-Score Data Analysis Teacher F with A same number of
Students

z-Test: Two Sample for Means

Variable 1 Variable 2

Mean 76.81481 61.48148

Known Variance 283.7064 278.7682

Observations 27 27

Hypothesized Mean Difference 0

Z 3.359441

P(Z<=z) one-tail 0.000391

z Critical one-tail 1.644854

P(Z<=z) two-tail 0.000781

z Critical two-tail 1.959964

As computed earlier, Z-score data Analysis ITC/ITP teacher “F” with other
teachers who teaches same subjects on the basis different number of students
has been computed. According to the summarized results, there is significant
performance among teacher “F” with all of the teachers.

As computed earlier, Z-score data Analysis ITC/ITP teacher “F” with other
teachers who teaches same subjects on the basis of same number of students
has been computed. According to the summarized results, there is significant
performance among teacher “F” with all of the teachers.

63
Table 4.24: Z-Score Data Analysis Teacher F with other different number
of Students

Teacher F Z Z Critical two-tail

A 4.1168 1.96
B 7.0827 1.96
C 3.0686 1.96
D 5.4561 1.96
E 3.1293 1.96
G 4.7683 1.96
H 3.8317 1.96
I 3.8953 1.96
J 3.4076 1.96
K 3.3785 1.96
L 3.4421 1.96

Table 4.25: Z-Score Data Analysis Teacher F with other same number of
Students

Teacher F Z z Critical two-tail

A 3.3594 1.96
B 5.6999 1.96
C 2.4668 1.96
D 5.2634 1.96
E 1.323 1.96
G 3.9718 1.96
H 2.4788 1.96
I 3.1042 1.96
J 2.8519 1.96
K 2.6697 1.96
L 2.3868 1.96

64
Summary
To prove our point, whether student’s performance or grades are dependent on
teacher’s performance or not, we have taken an example to demonstrate. We
have selected the course cal-1 to justify our analysis. For this task, we have
taken courses; Cal-I & Cal-II and ITC/ITP & CP/OOP. Using different
approaches for on the basis these results, we have ended with following
assumptions.

There is no significant difference between actual and predicted average of

student’s grades on the basis of teacher’s combinations we discussed.

When Cal-1 and ITC/ITP is taught by different teacher and get average marks
of students from session 141 to 163 in Cal-1 and ITC/ITP teacher-wise using
five different experiments, like : Average of actual Cal-I marks based on the
Cal-I teacher, Prediction of the Cal-I marks excluding the teacher name and
then taking the average teacher-wise, Prediction of the Cal-I marks including
the teacher and then taking the average teacher-wise , Average of actual Cal-II
marks based on the Cal-I teacher , Prediction of Cal-II marks and then
computed the average of these marks based on the Cal-I teacher, our predicted
results were higher than the actual results or nearly lesser than actual ones. The
result has showed that all teachers have different average in these subjects and
performances of teachers in this subject are significant.

For proved these result have used ANOVA analysis .The performance of
teacher in CAL-I and ITC/ITP have different and average of teachers have
significant between all teacher .The teacher whose performance is better might
have a reason that he is lenient in giving grades to the students and the teachers
whose performance is comparatively low might be strict in giving grades to the
students. Z-Score have proved these results and have showed that significant
result in CAL-I and ITC/ITP according to each teacher. With the help of these
result, we have recommended for CAL-I teacher E is a very suitable teacher
and ITC/ITP teacher F can improve the performance of students.

We have proposed a recommender system, which will recommend teachers to

the students based on their interest level and understanding and communication

65
with their teachers. With the help of recommender system, university
department may become able to identify the student’s performance and enable
them to make their performance better. To evaluate our prediction model and
recommender system, we have produced the answers of research questions. We
have found following strong observations during the course of our experiments.

 In context of CUST university Cal-I and ITC are critical subjects in

semester-l

 Recommendation of appropriate teacher for Cal-l is B and G

 Recommendation of appropriate teacher for OOP is E and F

66
Chapter 5

CONCLUSION AND FUTURE WORK

 Prediction is a helpful activity to reduce the attrition

 Existing factor is applicable to the certain level of accuracy and a set of

attributes is identified which is applicable in context of our university

 In context of CUST university Cal-I and ITC are critical subjects in

semester-l

 Recommendation of appropriate teacher for Cal-l and ITC

Our proposed research has been distributed in five chapters. First chapter
consist of introduction of the topic and purpose of this research. Second chapter
contains literature review in which the related work of the other researchers has
been presently in quite understandable manner. Third chapter contains
methodology and chapter no 4 contains detailed discussion of results and
experiments.

In fist chapter, we have briefly explained the background of our research topic
along with its scope, significance and applications. Moreover, we have
discussed research question, which we have constructed after critically
reviewing of literature. We have explored various data mining techniques that
were proposed and used by many researchers. The area of data mining is quite
versatile. Data mining means to explore intrinsic information from the bulk of
datasets. These datasets might belong to educational institutions or internet or
any organization. With the help of mining techniques such as classification, we
can make data analysis for prediction and find the accuracy. Researchers have
worked a lot in the field of education in data mining in which they have used
student’s academic record for specific reasons. They have used the records for
grade prediction, evaluation of student’s performance and teacher’s
performance most commonly.

67
In this regard, from literature we have obtained exiting factors such as
institutional factors and demographic factors. These factors are known as
exiting because they have applied in the western universities. We have applied
these factors in our local context. We have collected data from the Capital
University of Science and Technology, Pakistan. We have used the data of
BSCS program from six terms of spring, which covers the data of 3 years. In
this dataset, we have used the data of students of first semester and their final
grade has been predicted based on their midterm marks, term marks, gender,
age, Matric grades and inter grade. These attributes cover demographic,
institutional and pre-university attributes of the students. In addition,
researchers have used only two factors; Demographic and Pre-university. In our
experiments, we have considered another factor which is institutional in which
midterm and term marks are combined to predict the students final GPA/Grade.
This is how we have proposed prediction model on the basis students
associated attributes.

After discussion of literature review critically, we have identified different

classification approaches by the researches, so we used in our experimental
work as well. Afterwards, we have discussed chapter number 3 in which
diagram of proposed methodology have been mentioned and each step of
methodology has been discussed in detail. Our chapter number 4 contains
relative experiments and results that we have carried out through the help of
WEKA tool. WEKA is specifically designed to perform analysis in the field of
educational data mining. In the part of conclusion we have summarize the
results against our each research question.

We have proposed a recommender system, which will recommend teachers to

the students based on their interest level and understanding and communication
with their teachers. With the help of recommender system, university
department may become able to identify the student’s performance and enable
them to make their performance better. To evaluate our prediction model and
recommender system, we have produced the answers of our three research
questions. Those research questions and their answers have been summarized
as follows:

68
RQ1. Whether exiting identified factors for grade prediction are valid in our
local context (Pakistan)?
The answer to this research question is yes because we have gained quite
considerable accuracy from all data analysis measures. These analyses have
been applied on categorical and nominal data. Results in both categories are
same, which shows the validity of, existing factors in local context. All three
factors contributed well in order to increase the accuracy through WEKA tool
when ARPP format of data was given to it and classification algorithms; SMO
and Linear regression were applied.

RQ2. Which factors help in accurate prediction of student’s GPA of first

semester?
We have given answer to this research question in quite interesting fashion. We
have make combinations of all attributes that belong to these three categories of
factors. When we combined term marks and mid-term grades along with inter
and matriculation marks as well as with their gender and age; accuracy of
algorithms were boost up to the mark able level. Therefore the answer to this
research question is that demographic, pre-university and institutional factors
(GPA, Internal marks, #of Credits Hours, Program). Based on these results, we
have proposed a prediction model that will help to predict the final GPA/Grade
of the students and then can be evaluated in other departments of the university
too. There is not a single attribute that is affecting the performance of students
but a combination of attributes that also occurs in terms of hybrid approach.

RQ3. Is it possible to improve the performance of a student by allocating of

appropriate teacher for a subject? To answer the research question no 3, we have
drawn inferences by two ways. One way is statistical experiments and other
way is prediction experiments. Both types of experiments have sub experiments
associated with them. by statistical experiments, we have computed ANOVA
test on the group of attributes to extract the information of such teachers having
significant performance. But ANOVA has a limitation; as it is applied in group
of attributes; therefore we are unable to identify the teacher having significant
performance. To overcome this shortcoming of inferences, we have applied Z-
test through which we have obtained the teachers having significant

69
performance with the pairing of each teacher belong to that particular course.
By this way, we became able to bring out the teachers whose performance is
better than other teachers who teach same course. Then we computed overall
average performance of the teacher based on the overall average grade of
students whom they taught.
With the help of prediction experiments, we have drawn prediction of grades
inclusively and exclusively teacher’s name who taught the courses. Then
predicted results and actual results have been computed and compared. By this
way we became able to get the entire set of results in which our analysis shows
that predicted and actual results remained close at some point and better at
some point too. In the same time, predicted were bit lower than actual ones as
well which can be seen in chapter 4.

At the end conclusion and future work has been discussed.

5.1 Future Work

● The experimental work can be applied in other departments of the
university.
● This data set is relatively small. We can carry out same experiments in
other field of department and large datasets.
Moreover, these experiments can be carried out in government institution such
as NADRA.

70
 REFERENCES

Abaya, S. A., & Gerardo, B. D. (2013, September). An education data

mining tool for marketing based on C4. 5 classification technique. In e-
Learning and e-Technologies in Education (ICEEE), 2013 Second
International Conference on (pp. 289-293). IEEE.

Aher, S. B., & Lobo, L. M. R. J. (2012). A comparative study of association

rule algorithms for course recommender system in e-learning. International
Journal of Computer Applications, 39(1), 48-52.

Al-Barrak, M. A., & Al-Razgan, M. (2016). Predicting student’s final GPA

using decision trees: a case study. International Journal of Information and
Education Technology, 6(7), 528.

Anand, S. S., Bell, D. A., & Hughes, J. G. (1996). EDM: A general

framework for data mining based on evidence theory. Data & Knowledge
Engineering, 18(3), 189-223.

Azarcon Jr, D. E., Gallardo, C. D., Anacin, C. G., & Velasco, E. (2014).
Attrition and retention in higher education institution: A conjoint analysis
of consumer behavior in higher education. Asia Pacific Journal of
Education, Arts and SCIENCE, 1(5), 107-118.

Baker, R. S., & Inventado, P. S. (2014). Educational data mining and

learning analytics. In learning analytics (pp. 61-75). Springer New York.
Baradwaj, B. K., & Pal, S. (2012). Mining educational data to analyze
students' performance. arXiv preprint arXiv: 1201.3417.

Bydžovská, H. (2013). Course Enrolment Recommender System (Doctoral

dissertation, Masarykova univerzita, Fakulta informatiky).

71
Cabrera, A. F., Nora, A., & Castaneda, M. B. (1993). College persistence:
Structural equations modeling test of an integrated model of student
retention. The journal of higher education, 64(2), 123-139.

Dougiamas, M., & Taylor, P. (2003). Moodle: Using learning communities

to create an open source course management system.

Drea, C. (2004). Student Attrition and Retention in Ontario's

Colleges. College Quarterly, 7(2), n2.

Elbadrawy, A., Polyzou, A., Ren, Z., Sweeney, M., Karypis, G., &
Rangwala, H. (2016). Predicting Student Performance Using Personalized
Analytics. Computer, 49(4), 61-69.

Garcia-Saiz, D., & Zorrilla, M. E. (2011, November). Comparing

classification methods for predicting distance students' performance. In
WAPA (pp. 26-32).

Guarín, C. E. L., Guzmán, E. L., & González, F. A. (2015). A model to

predict low academic performance at a specific enrollment using data
mining. IEEE Revista Iberoamericana de Tecnologias del
Aprendizaje, 10(3), 119-125.

Hung, J. L., & Zhang, K. (2008). Revealing online learning behaviors and
activity patterns and making predictions with data mining techniques in
online teaching. MERLOT Journal of Online Learning and Teaching.

Inoue, S., Rodgers, P. A., Tennant, A., & Spencer, N. (2017). Reducing
Information to Stimulate Design Imagination. In Design Computing and
Cognition'16 (pp. 3-21). Springer, Cham.

72
Kabakchieva, D. (2013). Predicting student performance by using data
mining methods for classification. Cybernetics and Information
Technologies, 13(1), 61-72.

Kaminski, J. (2005). Moodle–a User-Friendly, Open Source Course

Management System. Online Journal of Nursing Informatics, 9(1).
Kaur, G., & Singh, W. (2016). Prediction Of Student Performance Using
Weka Tool.

Kizilcec, R. F., Pérez-Sanagustín, M., & Maldonado, J. J. (2017). Self-

regulated learning strategies predict learner behavior and goal attainment in
Massive Open Online Courses. Computers & Education, 104, 18-33.

Kokina, J., Pachamanova, D., & Corbett, A. (2017). The role of data
visualization and analytics in performance management: Guiding
entrepreneurial growth decisions. Journal of Accounting Education.
Kovacic, Z. (2010). Early prediction of student success: Mining students'
enrolment data.

Merceron, A., & Yacef, K. (2005, May). Educational Data Mining: a Case
Study. In AIED (pp. 467-474).

Miller, B. N., Konstan, J. A., & Riedl, J. (2004). PocketLens: Toward a

personal recommender system. ACM Transactions on Information Systems
(TOIS), 22(3), 437-476.

Nam, C. S., & Smith-Jackson, T. L. (2007). Web-based learning

environment: A theory-based design process for development and
evaluation. Journal of information technology education, 6.

Oskouei, R. J., &Askari, M. (2014).Predicting Academic Performance with

Applying Data Mining Techniques (Generalizing the results of two

73
Different Case Studies). Computer Engineering and Applications
Journal, 3(2), 79-88.

Peña-Ayala, A. (2014). Educational data mining: A survey and a data

mining-based analysis of recent works. Expert systems with
applications, 41(4), 1432-1462.

Polyzou, A., & Karypis, G. (2016). Grade prediction with models specific
to students and courses. International Journal of Data Science and
Analytics, 2(3-4), 159-171.

Ramist, L. (1981). College student attrition and retention. Romero, C., &
Ventura, S. (2007). Educational data mining: A survey from 1995 to 2005.
Expert systems with applications, 33(1), 135-146.

Romero, C., & Ventura, S. (2010). Educational data mining: a review of the
state of the art. IEEE Transactions on Systems, Man, and Cybernetics, Part
C (Applications and Reviews), 40(6), 601-618.

Romero, C., & Ventura, S. (2013). Data mining in education. Wiley

Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 3(1),
12-27.

Sachin, R. B., & Vijay, M. S. (2012, January). A survey and future vision
of data mining in educational field. In Advanced Computing &
Communication Technologies (ACCT), 2012 Second International
Conference on (pp. 96-100). IEEE.

Scheuer, O., & McLaren, B. M. (2012). Educational data mining.

In Encyclopedia of the SCIENCE of Learning (pp. 1075-1079). Springer
US.

74
Tewari, A. S., Saroj, A., & Barman, A. G. (2015). e-Learning
Recommender System for Teachers using Opinion Mining. In Information
Science and Applications (pp. 1021-1029). Springer Berlin Heidelberg.

Tinto, V. (1987). Leaving college: Rethinking the causes and cures of

student attrition. University of Chicago Press, 5801 S. Ellis Avenue,
Chicago, IL 60637.

Yang, D., Sinha, T., Adamson, D., & Rosé, C. P. (2013, December). Turn
on, tune in, drop out: Anticipating student dropouts in massive open online
courses. In Proceedings of the 2013 NIPS Data-driven education
workshop(Vol. 11, p. 14).

Zaíane, O. R. (2002, December). Building a recommender agent for e-

learning systems. In Computers in Education, 2002. Proceedings.
International Conference on (pp. 55-59). IEEE.

Stallone, M. N. (2011). Factors associated with student attrition and

retention in an educational leadership doctoral program. Journal of College
Teaching & Learning (TLC), 1(6).

Romero, C., & Ventura, S. (2013). Data mining in education. Wiley

Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 3(1),
12-27.

Zhang, Y., Oussena, S., Clark, T., & Hyensook, K. (2010). Using data
mining to improve student retention in HE: a case study.

Witten, I. H., Frank, E., Trigg, L. E., Hall, M. A., Holmes, G., &
Cunningham, S. J. (1999). Weka: Practical machine learning tools and
techniques with Java implementations.

75