International Journal of Scientific Research in Computer Science, Engineering and Information Technology

ISSN : 2456-3307 (www.ijsrcseit.com)

doi : https://doi.org/10.32628/CSEIT2172107
A Comparative Study on University Admission Predictions Using Machine
Learning Techniques
Prince Golden1, Kasturi Mojesh1, Lakshmi Madhavi Devarapalli1, Pabbidi Naga Suba Reddy1, Srigiri Rajesh1,
Ankita Chawla2
1Department of Computer Science and Engineering, Lovely Professional University, Phagwara, Punjab, India
2 Assistant Professor, Department of Computer Science and Engineering, Lovely Professional University,
Phagwara, Punjab, India

Article Info ABSTRACT

Volume 7, Issue 2 In this era of Cloud Computing and Machine Learning where every kind of work is
Page Number: 537-548 getting automated through machine learning techniques running off of cloud servers
to complete them more efficiently and quickly, what needs to be addressed is how
Publication Issue : we are changing our education systems and minimizing the troubles related to our
March-April-2021 education systems with all the advancements in technology. One of the the
prominent issues in front of students has always been their graduate admissions and
Article History the colleges they should apply to. It has always been difficult to decide as to which
university or college should they apply according to their marks obtained during
Accepted : 20 April 2021
their undergrad as not only it’s a tedious and time consuming thing to apply for
Published : 28 April 2021
number of universities at a single time but also expensive. Thus many machine
learning solutions have emerged in the recent years to tackle this problem and
provide various predictions, estimations and consultancies so that students can easily
make their decisions about applying to the universities with higher chances of
admission. In this paper, we review the machine learning techniques which are
prevalent and provide accurate predictions regarding university admissions. We
compare different regression models and machine learning methodologies such as,
Random Forest, Linear Regression, Stacked Ensemble Learning, Support Vector
Regression, Decision Trees, KNN(K-Nearest Neighbor) etc, used by other authors in
their works and try to reach on a conclusion as to which technique will provide
better accuracy.
Keywords - Linear Regression, Regression Models, Machine Learning Models for
prediction, Admission Predictions, Post Graduate studies, Prediction System, Data
Mining Techniques.

I. INTRODUCTION during their final year of their undergrad programme,

going through various tests such as GATE, CAT, GRE,
Undergrad students who wish to go for higher TOEFL, IELTS etc. Good scores in these exams with
education often start their preparations for their their GPA from their undergraduate programme
admissions in their dream universities and college insures their admissions in their dream colleges.

Copyright: © the author(s), publisher and licensee Technoscience Academy. This is an open-access article distributed under the 537
terms of the Creative Commons Attribution Non-Commercial License, which permits unrestricted non-commercial use,
distribution, and reproduction in any medium, provided the original work is properly cited
 Prince Golden et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, March-April-2021, 7 (2) : 537-548

Students want to pursue their graduation from best any student’s GPA, GRE results, LOR and SOP
universities and colleges with good rankings through the model and the model should predict the
worldwide. For that many choose to study abroad chances of the student’s admittance to the particular
and thus they take a slew of above mentioned tests university [3]. But the precision or accuracy of this
that makes it easier for them to apply and get selected model will not be that high to be implemented in real
for admissions in their preferred institutions. So the life scenarios as it’s based upon a very basic machine
graduate program takes immense preparation and it is learning algorithm. Thus to increase the accuracy of
one of the exhaustive tasks there is to apply and get the model various different machine learning
selected for one’s preferred institutions [1]. Hence, to techniques such as KNN, Random Forest, Support
lessen the troubles faced by the students and make Vector Regression, Decision trees etc have been used
their lives easier many efforts has been put through by researchers.
various researchers using ML techniques using
various regression models. Here in this paper, we will Thus in this paper, we will be focusing on to these
be reviewing all those works. We will also compare Machine Learning techniques in detail and try to
the methods which has been used by other authors to understand their functioning and efficiency and
render a model and try to find out which review them to find out the best technique so far for
methodology is the most efficient and by how much the accurate prediction of admission in any
using the data provided by them. As most of our university[1].
students try to attain their higher education abroad
and most likely from USA[2], we will be using the II. LITERATURE REVIEW
data set used in the universities of USA such as GRE,
GPA, SOP, LOR for the admission predictions which There has been many papers on the topic of
are readily available in the UCLA graduate database admission predictions but they all use varying
for admissions [1]. An efficient university admission methods using different machine learning approaches.
predictor will be very helpful as it will narrow down Few of them uses some format of UCLA dataset
the options available to a student for his/her which has the test scores which are more suitable for
application to any university as well as give them a US universities admission for the admission
certain idea of the chances of their admission in prediction for training the data model, others use the
particular university thus eliminating the need to academic merits, background and college admission
waste time and money on approaching different criteria etc to achieve the prediction. Lets have a look
consultancies and at the same time helps the students at various research works.
to be more informed about the situation [3].
[1] used the dataset form UCLA database and during
Basically the admission prediction can even be their research the data had been downloaded 2000+
achieved through simple Machine Learning times and the dataset had parameters which were
algorithm such as a simple Linear Regression model, significant as they were used by the admission
if for example we obtain the data such as GPA, GRE, committees for granting admission. The dataset
LOR, SOP etc for previous admission applications in included GRE, TOEFL, graduate GPA, Statement of
any particular university over the years and then Purpose (SOP) and Letter of Recommendation (LOR).
craft a Linear Regression model using the past data. Then the scrapped data is cleaned and formatted to
Now that we have that model we just need to pass suit the needs of an Indian student. After the

Volume 7, Issue 2, March-April-2021 | http://ijsrcseit.com 538

 Prince Golden et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, March-April-2021, 7 (2) : 537-548

segregation of the required data they have evaluated criteria. The prototype system then used live data of
their dataset with 4 different machine learning 2000+ high school student record and it suggests that
algorithms namely 1. Linear Regression 2. Support the hybrid model of decision trees and neural
Vector Regression 3. Random Forest and 4. Decision networks performs better than any model which use
Trees and after the evaluation and calculations of only decision trees or neural network and also
errors and tabulation of the results they have chosen provides flexibility as the prediction provides the
the best algorithm for their work. university based on the match of student’s profile.

There are other papers such as [3] which implement [2] Proposed a University Admission Predictor (UAP)
newer machine learning techniques such as Stacked system. They have tried to determine the predictions
Ensemble Learning. using various machine learning algorithms such as
The idea behind the application of this method KNN, Decision Trees, Ridge Regression etc to predict
dictates the training of several models and then the admission chances using the variables such as
combining their results to reach a solution for any GRE, TOEFL, GPA, SOP, LOR etc. During the
given problem. Traditionally, the ensemble approach research they have tried to compare the algorithms
was used to merge weak algorithms together as when mentioned above for choosing the best for their
put together they prove efficient. Thus a stacked proposed UAP system. The proposed system is able to
ensemble model is a model in which several machine predict the admission chances with 79% of average
learning algorithms are put together using meta- accuracy.
machine learning model, thus the predictions from
multiple sub-models become the features of the [4] used data mining techniques to predict the
meta-model increasing its efficiency. They collected performance of the students so that the predictions
data such as GRE scores, graduate GPA, TOEFL then can be used to support
scores, number of publications, their program of Decision making for university admissions using
application etc. In doing so a comparison with other Artificial Neural Networks (ANNs) and compared
supervised Machine Learning algorithms such as with other classification techniques such as Decision
Linear Regression, Decision Trees, Random Forest, Trees, Naive Bayes and Support Vector Regression.
Naive Bayes Classifier, Logistic Regression etc The study tried to fill the research gap that could
according to their efficiency, has also been done in help universities finding the predictions about the
this paper and an observation has been made that this academic performance of the students in the
approach is more efficient than many previous university prior to the admissions thus helping the
university admission prediction models. universities shortlist the students who will perform
better if given a chance at admissions and hence the
[6] have proposed a hybrid model using a university can gain better ratings as a result of better
combination of machine learning techniques to performing students. The system used the high school
tackle the student admission predictions. The system GPA, Scholastic Achievement Admission test score
proposed in the paper is built on top of a hybrid and General Aptitude Test score of 2039 students
model of decision trees and neural networks. Thus enrolled in computer science and information college
the proposed system predicts the chances of of a Saudi public college for the dataset. The research
admission in any specific university based on their also stated that the Scholastic Achievement
academic merits, background and college admission Admission test score was the criteria which made

Volume 7, Issue 2, March-April-2021 | http://ijsrcseit.com 539

 Prince Golden et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, March-April-2021, 7 (2) : 537-548

predictions with highest accuracy using Artificial [8] proposed a University Student Result Analysis
Neural Networks which has an accuracy rate above and Prediction system by implementing the Decision
79% and thus ANNs outperforms the other classifiers Tree machine learning algorithm. The system can be
such as Decision Trees, Naive Bayes and Support used to predict the student results by analyzing their
Vector Regression. previous year/semester marks and the model also
outlines the subjects in which the student is lacking.
[5] tried to come up with a machine learning The system tries to help students find out the area or
prediction system to predict the MS admission using the subject where a student needs to put more focus
the hypothetical UCLA data for the graduate so that he/she can avoid poor achievement and
admissions such as the GRE, TOEFL, GPA, SOP, LOR, perform better. The data used for the prediction
Extra curricular activities, outstanding achievements model is taken from the university database as well as
and projects and research done by the applicant. The through surveys done using google form. The data set
prediction system applied various machine learning contains information such as gender, extracurricular
algorithms such as Logistic Regression, Linear activities, number of tuition, programming skills,
Discriminant Analysis (LDA), KNN, Decision Tree, class test marks, assignment marks, attendance and
Random Forest, Naive Bayes and Support Vector previous GPA. The study tried to relate the
Machines (SVM) to find out the most accurate performance prediction closely with CGPA. To train
algorithm for the prediction. The author concludes the data model weka tools was used. The objective of
that Logistic Regression (LR) and Linear Discriminant the study also revolves around the comparison of
Analysis (LDA) performs best for this type of three different types of decision trees as well and
prediction with the dataset available and chose LR for concludes that the J48 decision tree algorithm
the prediction model. outperforms the other two algorithms achieving the
highest accuracy.
[7] proposed a Hybrid Recommender System for
college admission prediction using data mining [10] built a web based college admission predictor
techniques and knowledge discovery rules. The system to reduce the time taken for the seat
HRSPCA system used a cascade of two recommender allotments after the entrance exams. The system tried
algorithms for the predictions. The first to ease the problems of students as well as for the
recommender is used to determine the track of a college administration. The system eases up the
student for the preparatory exam and the second manual work needed for the seat allotments as well as
recommender assigns the specific college for the helps in the computerization of the same. The web
student if the student clears the preparatory exams. based system helps students decide their potential
Thus in this way the HRSCPA system achieves better future colleges and after shortlisting they can apply
performance. The college predictor algorithm uses to only those colleges where their chances of
historical student GPA data along with academic admission is more. The students need to register and
merits, background, student records and admission feed their academic as well as their personal
criteria of the colleges to predict the colleges. The information such as marks, background details etc
system was tested with live data of students from the and then accordingly the system will predict their
database of the King Abdulaziz University (KAU). potential colleges. As the predictions come based on
the entrance exams, the admin section is also
maintained so that the prediction also reflects to the

Volume 7, Issue 2, March-April-2021 | http://ijsrcseit.com 540

 Prince Golden et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, March-April-2021, 7 (2) : 537-548

admin and then after minimal manual verification USA. The proposed system is used to help the student
and with increased ease due to computerization of go for the graduate admissions in their preferred
the said subject the admin can allot the seat to the colleges based on their academic profile, hence it
student. Thus the system enhances the speed and does a matchmaking process and determines the best
flexibility of admissions. The system uses generic college or university for a student. The proposed
algorithms for the computation and the website made system used the means of database. It transformed the
for the system is made using PHP. The paper also relational database for all the specific and relevant
touched on the basics of software development and data of students into a universal database, I.e the
tried to explain the implementation of the proposed database after transformation had the academic
system, such as White and black box testing with information of the students who got successful
system and database design, modular design, input admissions in universities abroad. Then the proposed
and output design etc. system with the help of an algorithmic model
calculates similarities between the data of students
[9] proposed a model for any engineering students who got successful admissions and the student who
academic performance prediction using various wants to get the admission based on the data
Linear Regression techniques as well as generic data provided by the willing student I.e the similarity is
mining techniques, The regression techniques being check between the training data and the test
discussed in the paper are Naive Bayes, Decision Tree, data based on various weighted score using mean
Support Vector Machine (SVM), Neural Networks, squared deviation metrics. The proposed system used
KNN etc, but the paper used Linear Regression, KNN or K-nearest neighbour machine learning
Logistic Regression and Multivariate Linear algorithm for the calculation of top K universities for
Regression for its prediction model. The paper the N similar users. The system uses various
proposed to build a machine learning model to standardized test scores applicable in the USA such as
predict the academic performance of students of GRE, TOEFL, IELTS as well as the academic profile
engineering discipline.The model or the area of of the student such as their CGPA in undergrad from
research revolves around the prediction of any the customized relation database to train and test the
student’s performance in relation with his/her data set through the KNN algorithm to calculate the
amount of time spent on the internet. That is the similarities.Thus after the prediction the system lists
model tried to predict the outcomes based on the a number of universities the student can apply to for
independent variables such as usage of internet for admissions with a better chances of getting accepted
academic purposes, usage of internet for for admission.
communication purposes, usage of internet for
entertainment or social medial purposes, active [12] proposed a graduate school recommendation
duration on internet, usage of internet before system for easing the troubles of students as the
semester exams etc, with the output being the proposed system tries to help the students in decision
cumulative grade point average or CGPA being the making through shortlisting and recommending the
dependent variable throughout the research. appealing universities or colleges to the students
according to their needs and requirements. The
[11] presented an approach of designing and recommendation uses the Multi-Class Support Vector
developing a recommender system for the students Machine (SVM) and K-Nearest Neighbour machine
who wish to go for the graduate admissions in the learning approaches. For the recommendation system,

Volume 7, Issue 2, March-April-2021 | http://ijsrcseit.com 541

 Prince Golden et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, March-April-2021, 7 (2) : 537-548

the researchers have used SVM and KNN, SVM is very effective during computations of similarity
chosen for its high accuracy with the complex linear based upon the nearness of the objects or classifiers
and nonlinear data classification as the graduate data which are other colleges/universities in this case.
includes nonlinear type of data. With SVM, the core Thus through the collective use of SVM and KNN,
or the most appealing university is finalized based on this recommendation model promises better accuracy
the requirements of the student and the than other recommendation systems in the past. The
university/college through the fitness of classification. data set used for this system is taken from online
And after a core has been finalized, the KNN is used portals like Edulix.com [14] and Yocket.com [13]
to predict a number of universities/colleges that are which archive data of various universities.
similar to the core defined by the SVM as the KNN is

Table 1. Comparison of Algorithms based on Accuracy with respect to various research paper including
precision ratio
S.No
Algorithms/ Avg. Accuracy Precision F-measure Reference Paper
Methodologies used
1.
Linear Regression 72% [1], 72% [1] 72% [1] [1], [2]
79%[2]

2. Decision Tree 65% [3], 65% [3] 65% [3] [3], [4], [8]
83% [8],
78% [4]
3. Random Forest 66% [1],
77% [2], 63% [3] 63% [3] [1], [2], [3]
62.5%[3]
4. Support Vector Machines 64% [1], 55% [3] 55% [3] [1], [3], [4]
(SVM) 55% [3],
72% [4]
5. Naive Bayes Classifier 53.3% [3], 54%[3] 47% [3] [3], [4]
73% [4]
6. Stacked Ensemble 91% [3] 91% [3] 91% [3] [3]
Learning
7. Ridge Regression 78% [2] NA [2] NA [2] [2]

8. Artificial Neural 79.22% [4], 56% [3] 56% [3] [4],[3]

Networks (ANN) 56% [3]
9.
K-Nearest Neighbour 72$ [2], 57% [3] 57% [3] [3]

Volume 7, Issue 2, March-April-2021 | http://ijsrcseit.com 542

 Prince Golden et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, March-April-2021, 7 (2) : 537-548

(KNN) 57% [3]

10. Logistic Regression (LR) 55.88% 54%[3] 54%[3] [3]
[3]
11. Linear Discriminant 56% [3], 56% [3] 56% [3] [3], [5]
Analysis (LDA) 72% [5]
12. Hybrid Model Neural 80-90% 80-90% 80-90% [6]
Networks + Decision Tree [6] [6] [6]
13. Hybrid Model 55 % [12] 54% [12] 50% [12] [12]
SVM + KNN

III. METHODOLOGIES variable X independent. Thus, e.g. after a linear

regression, the weight (dependent variable) of an
Several machine learning algorithms have been used individual from their height may be estimated
by several authors in the above mentioned literature (independent variable)[16].
review. In this section we will be seeing those
Logistic Regression
methodologies and how these classifiers works. This
section also deals with the inner workings of the Logistic regression is similar to linear regression but
listed methodologies. instead of a linear response given by the linear
regression, the logistic regression has a binomial
Linear Regression
response variable. In the logistic regression we can
Linear regression is a regression model which uses have more than 2 continuous explanatory variables
supervised learning approach of machine learning. As and it’s easier to handle those variables
the name suggests, a regression task is performed simultaneously. In the case of more than one
with this algorithm and a model is created based on explanatory variable, logistic regression is used to
the regression to derive meaningful contexts from a calculate the odds ratio. With the fact that the
given set of data. It is one of the basic classifier used outcome variable is binomial, the technique is very
in machine learning. The algorithm is used to find similar to multiple linear regression. The effect of
out the relationship between dependent and each vector on the odds ratio of the observed
independent variables. It is used to predict a occurrence of interest is the result. The biggest
dependent value or dependent variable (y) based on a benefit is that when considering the relationship of
given independent variable (x) thus finding a linear all factors together, misleading consequences are
relationship between these two variables and thus avoided. A machine learning model created using
named as Linear Regression[16,15].The linear logistic regression models the chance of all the
regression model defines a variable which is based on outcomes based on the characteristics or features
a straight line described by equation Y = a +b x when provided to the model through data feeding [17].
one of the y-intersections of the line is y, and b is its
Decision Tree
pitch. First of all, the values of dependent variable Y
and independent variable X are calculated with the The decision tree is the most effective and widely
aid of statistical methods by parameters A and B of used classification and prediction method. A Decision
the regression line. The regression line allows one to tree is a tree structure that looks like a flowchart,
forecast the value of variable Y based on the value of with each internal node representing a test on an

Volume 7, Issue 2, March-April-2021 | http://ijsrcseit.com 543

 Prince Golden et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, March-April-2021, 7 (2) : 537-548

attribute, each branch representing the test's result, Find the best hyperplane to isolate the results. As a
and each leaf node (terminal node) holding a class result, SVM can only do classification task (i.e.,
name[15]. Decision trees are fundamentally a basic choose between two classes). Even so, there are
form of classifier, which is one of their several methods for dealing with multi-class
advantages[18]. In its most basic form, we ask yes-or- problems[15].
no questions, and each internal node has a 'yes' and a
Naive Bayes Classifier
'no' child. Following the model from the topmost
node, the root, to a node without children, a leaf, an The Naive Bayes classifier consist of a wide collection
item is sorted into a class depending on the responses of algorithms which are based on Bayes’ Theorem.
that refer to the item in review[18]. Since they Thus it’s not a single algorithm but a family of
combine basic questions about the data in an intuitive algorithms in which all of them share a common
manner, decision trees are often more easy to principle, i.e every pair of a selected feature or
interpret than other classifiers such as neural characteristic is classified as independent of each
networks and support vector machines. Methods for other. That makes this algorithm a simple learning
removing decision rules from decision trees have also algorithm utilizing Bayes’ rule with an assumption of
shown to be effective. Unfortunately, minor all the attributes being conditionally independent of
differences in input data will also result in significant the class. Although in practice the independence
changes in the tree's structure. Decision trees are assumption is often overlooked and violated, but still
adaptable enough to deal with objects that have a the classifier delivers competitive classification
combination of real-valued and categorical features, accuracy [15,20]. Nave Bayes is a technique for
as well as items that lack any features[18]. Thus estimating the posterior likelihood P(y | x) of each
decision trees can produce rules that are easy to class y given an object x using knowledge from
understand. Decision trees are used to build models sample data. We may use those projections for
without involving a lot of calculation and can grouping or other decision support implementations
accommodate factors that are both constant and until we have them[20]. The basic principle of Naive
categorical clearly showing which areas are most Bayes is that each function renders an independent
relevant for forecasting or classification[15]. and equivalent contribution to the result[15]. In
mathematical terms, Bayes theorem find the
Support Vector Machines (SVM)
probability of any given event’s occurrence with
Support Vector Machines (SVMs) are a family of respect to the probability of another event’s
supervised learning methods that can be used to solve occurrence.
regression and classification tasks. It is most often
K-Nearest Neighbour (KNN)
used for sorting, but it can also be very useful for
regression. A SVM classifier constructs a maximum- K-Nearest Neighbors is a simple but important
margin hyperplane in a transformed input space and classification technique in Machine Learning. It is a
divides the sample classes thereby minimizing the supervised learning algorithm that is widely used in
distance to the closest cleanly separated instances. pattern recognition, data processing, and intrusion
This hyper-plane is nothing other than a line in two detection. It is easily applicable in real-world
dimensions[15,19]. Each data object in the dataset is situations since it is non-parametric, which means it
plotted in an N-dimensional space in SVM, where N has no underlying assumptions regarding data
represents the number of characteristics in the data. delivery (as opposed to other algorithms such as

Volume 7, Issue 2, March-April-2021 | http://ijsrcseit.com 544

 Prince Golden et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, March-April-2021, 7 (2) : 537-548

GMM, which assume a Gaussian distribution of the mistakes (as long as they don't all err in the same
given data)[15,21]. The KNN algorithm presumes that direction all of the time). While certain trees will be
identical objects happen nearby. In other terms, incorrect, several more would be accurate, allowing
related objects are close together. The KNN algorithm the trees to migrate in the same direction as a group.
is based on the premise that this assumption is valid Thus a large number of relatively uncorrelated
for the algorithm to be useful. KNN captures the models (trees) acting as a committee will outperform
concept of similarity (also known as space, proximity, all of the constituent models individually[22].
or closeness) with some mathematics we might have
Artificial Neural Networks (ANN)
studied as children, such as measuring the distance
between points on a line. Thus KNN works by Artificial Neural Networks (ANN) are influenced by
calculating the distances between a query and all of the way biological neural systems interpret
the examples in the results, then choosing the information, such as the brain which is built of very
number of examples (K) nearest to the query and complex webs of interconnected neurons. The
voting for the most common symbol (in the case of information processing system is made up of several
classification) or averaging the labels (in the case of deeply integrated processing components (neurons)
regression)[21]. It's simple to set up and grasp, but it that collaborate to solve particular problems[23].
has a big downside in being slightly slower as the ANNs, just like people, learn by example[23]. The
scale of the data in use increases[15]. human brain is a tightly linked network of
approximately 10^11-10^12 neurons, with each
Random Forest
connected neuron linked to an average of l0^4-10^5
A Random Forest is an ensemble learning method other neurons. So, on average, it takes about 10-1
that can execute both regression and classification for a human brain to make remarkably complicated
tasks by combining several decision trees and a choices. The aim of ANN models is to catch this
technique known as Bootstrap and Aggregation, also kind of highly parallel computing based on
known as bagging. The basic principle is to use distributed representations. ANNs are typically
several decision trees to determine the final constructed from a densely interconnected series of
production rather than relying on individual decision simple units, each of which takes a number of real-
trees. Random Forest's foundation learning structures valued inputs and generates a single real-valued
are multiple decision trees. We randomly select rows output[15]..ANN learning, as learning in biological
and features from the dataset to create sample systems, requires changes to the synaptic
datasets for each model. This section is known as interactions that occur between neurons[23]. Thus
Bootstrap[15]. So it is made up of a large number of the electrical function of the brain and nervous
independent decision trees that work together as an system is simulated by ANN models. Processing
ensemble. Each individual tree in the random forest components (also known as neurodes or perceptrons)
produces a class prediction, and the class with the communicate with one another. The neurodes are
most votes becomes the prediction of a model. The typically organized in a layer or vector, with the
main point is the low correlation between models. output of one layer acting as the entry to the next
Uncorrelated models can generate ensemble layer and potentially other layers. A neurode can be
predictions that are more reliable than individual bound to all or a subset of the neurodes in the
predictions. The explanation for this great result is corresponding layer, simulating synaptic
that the trees shield each other from their individual interactions in the brain. Weighted data signals

Volume 7, Issue 2, March-April-2021 | http://ijsrcseit.com 545

 Prince Golden et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, March-April-2021, 7 (2) : 537-548

reaching a neurode mimic the electrical stimulation with high variances. The boost constructs some
of a neuron and, as a result, information flow within exponential versus small variance models to
the network or brain. The input data to a processing decrease the inclination[15,3]. Ensemble Learning is
unit, in, are amplified by a connection weight, a form of machine learning technique, which
W(n,m), which imitates neural pathway increasing training many weak models and combining the
in the brain. Learning is emulated in ANNs by results with them to solve a specific problem[14].
adjusting the link strength or weights[23,24]. An MLP alone is not the most effective modell of
the University Prediction Scheme neural network.
Linear Discriminant Analysis (LDA)
There are a couple of more sophisticated profound
Linear Discriminant Analysis (LDA) is a common learning processes. However, the results are
pre-processing technique for dimensionality extremely strong when you want to stack several
reduction problems in machine learning and pattern MLP models to build a stacked ensemble model. To
classification applications[25]. Since logistic do so, we start by training the submodels of the
regression is a classification algorithm, it has MLP and then use the predictions of the submodels
traditionally been restricted to two-class to train a metamodel. The outcome is a superior
classification problems. Linear Discriminant mode of deep learning[3]. Ensemble learning as a
Analysis is the favoured linear classification method whole consists of the strategical combination of
where there are more than two groups. It is used to various learning models (such as supervised
model group distinctions, such as dividing two or classification techniques) to create a superior model
more classes. It is used to project functions from a that can improve classification or estimation or even
higher dimension space to a lower dimension decrease opportunities to choose a bad outcome [3].
space[15]. We have two groups, for example, and Ensemble learning model has 2 types. The first
they need to be effectively separated. Multiple being Model Averaging Ensemble and second is
features can be included in classes. It can result in Stacked Generalization. In model averaging, the
any overlap using only one function to distinguish projection better results, the forecasts of different
them, as seen in the figure below. Therefore, we models are merged. This may also be regarded as the
will continue to increase the number of disadvantage of the model, since there may be a case
classification features[25,15]. To achieve maximum in which one variant is better than another but both
distance between means of two different classes and weigh same. Hence, each model is allocated weights
minimizing the variation within each class. The to address this situation. The best performing model
broad applications of LDA include Face Recognition, is weighing than a poorer model by adding weights
Medical classification of disease state as mild, severe, to each model. Finally, an average weighted and the
moderate based on diagnosis, Customer Type last projection is estimated[3]. Whereas in stacked
Identification etc. generalization or stacking the results of different
classification algorithms are combined so that to
Stacked Ensemble Learning
make a clearer prediction with the help of meta
Stacking is a means of grouping various classifier which is used to combine several models
classifications or models of regression. The most and at the end helps in generating the predictions
popular styles are Bagging or Boosting. There are through stacked ensemble learning model[3].
several ways of ensembles. Bagging enables the
average variation to decrease many related models

Volume 7, Issue 2, March-April-2021 | http://ijsrcseit.com 546

 Prince Golden et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, March-April-2021, 7 (2) : 537-548

IV. CONCLUSION V. REFERENCES

[1]. M. S. Acharya, A. Armaan and A. S. Antony, "A

The numbers of the students seeking further higher
Comparison of Regression Models for Prediction of
education abroad keeps increasing every year. And Graduate Admissions," 2019 International
this number won’t go down in upcoming years either. Conference on Computational Intelligence in Data
In our technological and competitive world the Science (ICCIDS), Chennai, India, 2019, pp. 1-5,
students constantly feel the need to upskill doi: 10.1109/ICCIDS.2019.8862140.
themselves and fare better than their peers and [2]. Chithra Apoorva D A, Malepati ChanduNath, Peta
competitors everyday and thus after their under Rohith, Bindu Shree.S, Swaroop.S, “Prediction for
graduation, a vast number of students tend to leave University Admission using Machine Learning”
International Journal of Recent Technology and
for higher education abroad. To facilitate an easy
Engineering (IJRTE) ISSN: 2277-3878, Volume-8,
migration, and to provide them with a tool which can
Issue-6 March 2020
help them shortlist their potential college/university
[3]. S. Sridhar, S. Mootha and S. Kolagati, "A University
based on various parameters will not just empower Admission Prediction System using Stacked
them in finding their preferred institutions but also Ensemble Learning," 2020 Advanced Computing
ensure that it saves their precious time as well as and Communication Technologies for High
their hard earned money at the same time. Many Performance Applications (ACCTHPA), Cochin,
researchers have tried to come up with ideas using India, 2020, pp. 162-167, doi:
machine learning, data mining, and generic 10.1109/ACCTHPA49271.2020.9213205.
algorithm from time to time, each with their own [4]. H. A. Mengash, "Using Data Mining Techniques to
Predict Student Performance to Support Decision
benefits and drawbacks. If we go through the
Making in University Admission Systems," in IEEE
research done till date, the success rate of hybrid
Access, vol. 8, pp. 55462-55470, 2020, doi:
systems with the implementations of more than one
10.1109/ACCESS.2020.2981905.
algorithm concurrently is higher with low error rates, [5]. Naman Doshi, “Predicting MS Admission”,
as the work done by each algorithm in any hybrid https://medium.com/data-science-weekly-
system relies on each of them only for the specific dsw/predictingms-admission-afbad9c5c599
task and hence later collaboration ensures higher February, 2018.
accuracy. In this regard, the stacked ensemble [6]. S. Fong, Y. Si and R. P. Biuk-Aghai, "Applying a
approach, which itself is based on an ensemble hybrid model of neural network and decision tree
approach which tries to bring different algorithms classifier for predicting university admission", 2009
7th International Conference on Information,
together to gain better accuracy, fares better than all
Communications and Signal Processing (ICICS),
the other approaches discussed and reviewed in this
Macau, pp. 1-5,2009
paper. An improvement can also be made and the
[7]. A. H. M. Ragab, A. F. S. Mashat and A. M. Khedra,
system can always be made more efficient and "HRSPCA: Hybrid recommender system for
accurate using NLP or natural language processing predicting college admission," 2012 12th
methods which can help rate the quality of the International Conference on Intelligent Systems
written essays and paragraphs during the exams, so Design and Applications (ISDA), Kochi, India, 2012,
that those features can also help narrow down the pp. 107-113, doi: 10.1109/ISDA.2012.6416521.
chances of admissions over time. [8]. Md. Imdadul Hoque, Abul kalam Azad, Mohammad
Abu Hurayra Tuhin, Zayed Us Salehin,"University
Students Result Analysis and Prediction System by

Volume 7, Issue 2, March-April-2021 | http://ijsrcseit.com 547

 Prince Golden et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol, March-April-2021, 7 (2) : 537-548

Decision Tree Algorithm" Advances in Science, [19]. Shmilovici A. (2005) Support Vector Machines. In:
Technology and Engineering Systems Journal Vol. Maimon O., Rokach L. (eds) Data Mining and
5, No. 3, 115-122 (2020) ,DOI: 10.25046/aj050315 Knowledge Discovery Handbook. Springer, Boston,
[9]. R.R.Rajalaxmi , P.Natesan , N.Krishnamoorthy MA. https://doi.org/10.1007/0-387-25465-X_12
,S.Ponni,"Regression Model for Predicting [20]. Webb G.I. (2011) Naïve Bayes. In: Sammut C.,
Engineering Students Academic Performance" Webb G.I. (eds) Encyclopedia of Machine Learning.
International Journal of Recent Technology and Springer, Boston, MA. https://doi.org/10.1007/978-
Engineering (IJRTE) ISSN: 2277-3878, Volume-7 0-387-30164-8_576
Issue-6S3 April, 2019 [21]. Zhang, Zhongheng. “Introduction to machine
[10]. Annam Mallikharjuna Roa , Nagineni Dharani , A. learning: k-nearest neighbors.” Annals of
Satya Raghava , J. Buvanambigai , K. translational medicine vol. 4,11 (2016): 218.
Sathish,"College Admission Predictor" Journal of doi:10.21037/atm.2016.03.37
Network Communications and Emerging [22]. https://towardsdatascience.com/understanding-
Technologies (JNCET) www.jncet.org Volume 8, random-forest-58381e0602d2
Issue 4, April (2018) [23]. Grosan C., Abraham A. (2011) Artificial Neural
[11]. M. Hasan, S. Ahmed, D. M. Abdullah and M. S. Networks. In: Intelligent Systems. Intelligent
Rahman, "Graduate school recommender system: Systems Reference Library, vol 17. Springer, Berlin,
Assisting admission seekers to apply for graduate Heidelberg. https://doi.org/10.1007/978-3-642-
studies in appropriate graduate schools," 2016 5th 21004-4_12
International Conference on Informatics, [24]. https://www.sciencedirect.com/topics/engineering/a
Electronics and Vision (ICIEV), Dhaka, Bangladesh, rtificial-neural-network
2016, pp. 502-507, doi: [25]. Tharwat, Alaa & Gaber, Tarek & Ibrahim,
10.1109/ICIEV.2016.7760053. Abdelhameed & Hassanien, Aboul Ella. (2017).
[12]. A. Baskota and Y. Ng, "A Graduate School Linear discriminant analysis: A detailed tutorial. Ai
Recommendation System Using the Multi-Class Communications. 30. 169-190,. 10.3233/AIC-
Support Vector Machine and KNN Approaches," 170729.
2018 IEEE International Conference on Information
Reuse and Integration (IRI), Salt Lake City, UT, Cite this article as :
USA, 2018, pp. 277-284, doi: Prince Golden, Kasturi Mojesh, Lakshmi Madhavi
10.1109/IRI.2018.00050.
Devarapalli, Pabbidi Naga Suba Reddy, Srigiri Rajesh,
[13]. Yocket,com
Ankita Chawla, "A Comparative Study on University
[14]. Edulix.com
Admission Predictions Using Machine Learning
[15]. Geeksforgeeks.com
[16]. Schneider, Astrid et al. “Linear regression analysis:
Techniques", International Journal of Scientific
part 14 of a series on evaluation of scientific Research in Computer Science, Engineering and
publications.” Deutsches Arzteblatt international Information Technology (IJSRCSEIT), ISSN : 2456-
vol. 107,44 (2010): 776-82. 3307, Volume 7 Issue 2, pp. 537-548, March-April
doi:10.3238/arztebl.2010.0776 2021. Available at
[17]. Sperandei, Sandro. “Understanding logistic doi : https://doi.org/10.32628/CSEIT2172107
regression analysis.” Biochemia medica vol. 24,1 12- Journal URL : https://ijsrcseit.com/CSEIT2172107
8. 15 Feb. 2014, doi:10.11613/BM.2014.003
[18]. Kingsford, Carl, and Steven L Salzberg. “What are
decision trees?.” Nature biotechnology vol. 26,9
(2008): 1011-3. doi:10.1038/nbt0908-1011

Volume 7, Issue 2, March-April-2021 | http://ijsrcseit.com 548