International Journal of Engineering Research & Technology (IJERT)

ISSN: 2278-0181
Vol. 4 Issue 03, March-2015

Student Performance Prediction using Machine

Learning
Havan Agrawal, Harshil Mavani
Department of Information Technology
K. J. Somaiya College of Engineering
Mumbai, India

Abstract - In this paper, a model is proposed to predict the

performance of students in an academic organization. The II. BACKGROUND AND RELATED WORK
algorithm employed is a machine learning technique called
Neural Networks. Further, the importance of several different A. Algorithm
attributes, or "features" is considered, in order to determine Classification is one of the most frequently studied
which of these are correlated with student performance. Finally, problems by data mining and machine learning (ML)
the results of an experiment follow, showcasing the power of researchers. It consists of predicting the value of a
machine learning in such an application.
(categorical) attribute (the class) based on the values of other
Keywords— Artificial intelligence, machine learning, student attributes (the predicting attributes). There are different
performance, neural networks classification methods.
Bayesian classification is an algorithm that is based on
I. INTRODUCTION Bayes rule of conditional probability. Bayes rule is a
There are many studies in the learning field that technique to estimate the likelihood of a property given the
investigated the ways of applying machine learning set of data as evidence or input. Bayes rule or Bayes theorem
techniques for various educational purposes. One of the is-
focuses of these studies is to identify high-risk students, as
well as to identify features which affect the performance of
students.
The study conducted by Kotsiantis et al [1] is one of the A more recent development in classification is that of
initial studies which investigated application of machine artificial neural networks. These networks are modeled after
learning techniques in distance learning for dropout the human neural system (hence the name), and have proven
prediction. The most significant contribution by this study to be as powerful, if not more, as any other algorithm. While
was that it was a pioneer and carved the path for several such implementations may be complex, these networks are capable
studies. While machine learning algorithms had been of understanding non-linear patterns in data.
previously implemented in several settings, this was perhaps A detailed description of the algorithms can be found in
the first time that these techniques were applied to an [4].
academic environment.
Kotsiantis et al [1] compared five algorithms, viz.
Bhardwaj and Pal [2] conducted a study in India, Decision Trees (C4.5), Naive Bayes algorithm (Bayesian
Faizabad to determine factors that most heavily affected networks), 3-NN (kNN), RIPPER (Rule Learning) and
student performance. They used Bayesian Classification for WINNOW (Perceptron based neural networks). This study
their study. was composed of two experimental stages, training and
The study by Erkan Er [3] was based upon Kotsiantis' as testing. During these stages, number of attributes was
well as other similar studies. It concluded that Naive Bayes increased step-by-step. For example, while only demographic
indeed performed better than any other machine learning data was included in the first step, performance attributes
algorithm. However, the crucial contribution of this study were added in the next step. Five algorithms were tested for
was that time-invariant features may be detrimental to the each these subsequent steps and then they were compared.
machine learning process, and hence are better left out of the This comparative study helped in narrowing down candidates
study entirely. He also concluded that "Instead of for our own application.
demographic characteristics of students, using initial However, classification of data into binary groups
attendance and homework grades produces better prediction seems insufficient. The primary goal of this study was only
rate at earlier stages." detecting at-risk students instead of determining performance
levels of students. Classifying students according to their
performances in different levels (e.g. poor, average, good,
excellent, etc.) might be more useful. In this way, instructors
can provide more adaptive feedback for each student.

IJERTV4IS030127 www.ijert.org 111

(This work is licensed under a Creative Commons Attribution 4.0 International License.)
 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
Vol. 4 Issue 03, March-2015

B. Features of this student. Octave was used for test purposes. The marks
Bhardwaj and Pal [2] conducted a study on the student of 80 B.E. I.T (Bachelor of Engineering, Information
performance based by selecting 300 students from 5 different Technology) students from semester 3 to semester 6 were
degree college conducting BCA (Bachelor of Computer used. The algorithm is trained on a training set of 60 students,
Application) course of Dr. R. M. L. Awadh University, and tested on a cross-validation set of 10 students, to predict
Faizabad, India. By means of Bayesian classification method marks in 6 subjects. This is done 7 times, varying the training
on 17 attributes, it was found that the factors like students' and test sets each time (k-fold cross validation). An error of
grade in senior secondary exam, living location, medium of plus or minus 8 marks was considered as accurate. The error
teaching, mother's qualification, students other habit, family statistics were as follows:
annual income and student's family status were highly Average error = 6
correlated with the student academic performance. Accurate = 296
In the present study, those variables whose probability Erroneous = 124
values were greater than 0.70 were given due considerations Accuracy Rate = 70.48%
and the highly influencing variables with high probability Once it was confirmed that the data conforms well to a
values have been shown in Table 1. These features were used machine learning algorithm, we conducted a comparative
for prediction model construction. For both variable selection study of neural networks and Bayesian classification, on the
and prediction model construction, the publishers have used basis of varying training and test sets. The results were fairly
MATLAB. surprising. In general, the neural networks tend to outperform
From the table, it is found that the second high potential Bayesian classification. This is somewhat justified once one
variable for students' performance is their living location, and realizes that the input provided to the algorithm was on a
the third high potential variable for students' performance is continuous range, and Bayesian classification traditionally
medium of teaching. In Uttar Pradesh the mother tongue requires discrete data.
language of students is Hindi. Hence, students tend to be Finally, an application was made that employed neural
more comfortable in Hindi and other languages, than in the networks (Figure 2). The application provides to and fro
English language. access of data from .csv (Comma Separated Values) files.
C. Uniqueness When a prediction is required, it dynamically trains a network
of 3 layers, and provided prediction of marks in discrete
The study conducted by Erkan Er [3] proved valuable in
classes of 20 marks.
confirming the uniqueness of the proposed application. His
work concluded that all current applications of machine The training dataset size was increased in increments of
learning in an academic setting were to predict dropout rates 10, starting from 40, for 17 subjects. The test set was of 10
in a distance learning program. There is perhaps no students, to predict a single subject. The accuracy results are
application that attempts to predict the absolute performance summarized in Table 2.
of the student. If one does exist, it has not been published yet.
D. Inference
We analyzed the experiments and results of the
aforementioned studies, and two prominent inferences were
drawn. The first is that Naive Bayes Classification proves to
be an excellent algorithm for the application of predicting
student performance in an academic setting. Further, a worthy
contender for the same is neural networks. Secondly, several
factors contribute to a student's performance, apart from
previous academic performance.

Table 1: Study Results

Variable Description Probability
GSS Student’s Grade in Secondary 0.8642
Education
LLoc Living Location 0.7862
Figure 1: Screenshot of Application
Med Medium of Teaching 0.7225
Table 2: Neural Network Accuracy
III. EXPERIMENTAL WORK Training Dataset Size Accuracy
Initially, the existence of a linear relationship between a 40 50 %
student's previous academic performance levels was 50 50 %
considered. This relationship can be expressed accurately 60 60 %
70 70 %
using Multivariate Linear Regression. Multivariate Linear
Regression uses past semester marks of a student and marks
scored by this student's senior batches to predict future marks

IJERTV4IS030127 www.ijert.org 112

(This work is licensed under a Creative Commons Attribution 4.0 International License.)
 International Journal of Engineering Research & Technology (IJERT)
ISSN: 2278-0181
Vol. 4 Issue 03, March-2015

IV. CONCLUSION REFERENCES

Present studies shows that academic performances of the
students are primarily dependent on their past performances. [1] S. Kotsiantis, C. Pierrakeas, and P. Pintelas, “Preventing student
dropout in distance learning systems using machine learning
Our investigation confirms that past performances have techniques,” AI Techniques in Web-Based Educational Systems at
indeed got a significant influence over students' performance. Seventh International Conference on Knowledge-Based Intelligent
Further, we confirmed that the performance of neural Information & Engineering Systems, pp. 3-5, September 2003.
networks increases with increase in dataset size. [2] B.K. Bharadwaj and S. Pal. "Data Mining: A prediction for
performance improvement using classification", International Journal
Machine learning has come far from its nascent stages, of Computer Science and Information Security (IJCSIS), Vol. 9, No. 4,
and can prove to be a powerful tool in academia. In the pp. 136-140, 2011.
future, applications similar to the one developed, as well as [3] Erkan Er. "Identifying At-Risk Students Using Machine Learning
Techniques", International Journal of Machine Learning and
any improvements thereof may become an integrated part of Computing, Vol. 2, No. 4, pp. August 2012.
every academic institution. [4] S. Kotsiantis, I.D. Zaharakis, and P. Pintelas, "Assessing Supervised
Machine Learning Techniques for Predicting Student Learning
ACKNOWLEDGMENTS Preferences"

We thank Prof. Yogita Borse. Without her guidance, this

paper could never have been accomplished.

IJERTV4IS030127 www.ijert.org 113

(This work is licensed under a Creative Commons Attribution 4.0 International License.)