See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/348433004

Graduate Admission Prediction Using Machine Learning

Article · December 2020

DOI: 10.46300/91013.2020.14.13

CITATIONS READS
0 3,162

4 authors, including:

Ali Bou Nassif Ismail Shahin

University of Sharjah University of Sharjah
133 PUBLICATIONS   2,245 CITATIONS    100 PUBLICATIONS   1,087 CITATIONS   

SEE PROFILE SEE PROFILE

Ashraf M Elnagar
University of Sharjah
110 PUBLICATIONS   1,099 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Emirati-Accented Speaker and Emotion Recognition Based on Deep Neural Network View project

Machine learning View project

All content following this page was uploaded by Ali Bou Nassif on 13 January 2021.

The user has requested enhancement of the downloaded file.

 Graduate Admission Prediction Using Machine
Learning
Sara Aljasmi Ali Bou Nassif Ismail Shahin Ashraf Elnagar
Department of Computer Department of Computer Department of Electrical Department of Computer Science
Engineering Engineering Engineering University of Sharjah
University of Sharjah University of Sharjah University of Sharjah Sharjah, UAE
Sharjah, UAE Sharjah, UAE Sharjah, UAE [email protected]
[email protected] [email protected] [email protected]

Abstract— Student admission problem is very important in • Statement of purpose (SOP) which is a document written to
educational institutions. This paper addresses machine learning show the candidate's life, ambitious and the motivations for
models to predict the chance of a student to be admitted to a the chosen degree/ university. The score will be out of 5
master’s program. This will assist students to know in advance if points.
they have a chance to get accepted. The machine learning models
are multiple linear regression, k-nearest neighbor, random forest, • Letter of Recommendation Strength (LOR) which verifies
and Multilayer Perceptron. Experiments show that the Multilayer the candidate professional experience, builds credibility,
Perceptron model surpasses other models. boosts confidence and ensures your competency. The score
is out of 5 points
Keywords— K- Nearest neighbors, Multilayer Perceptron, Multiple
linear regression, Random forest, Student admission • Undergraduate GPA (CGPA) out of 10

I. INTRODUCTION • Research Experience that can support the application, such

as publishing research papers in conferences, working as
The world markets are developing rapidly and continuously research assistant with university professor (either 0 or 1).
looking for the best knowledge and experience among people.
Young workers who want to stand out in their jobs are always One dependent variable can be predicted which is chance of
looking for higher degrees that can help them in improving their admission, that is according to the input given will be ranging
skills and knowledge. As a result, the number of students from 0 to 1.
applying for graduate studies has increased in the last decade
It is worth to mention that all tests will be done using R
[1]–[4]. This fact has motivated us to study the grades of
language. Models will be created using Weka, and statistical test
students and the possibility of admission for master’s programs
will be performed using PHStat.
that can help universities in predicting the possibility of
accepting master’s students submitting each year and provide II. TECHNICAL BACKGROUND
the needed resources.
A. Shapiro-Wilk Normality Test
The dataset [5] presented in this paper is related to
educational domain. Admission is a dataset with 500 rows that The Shapiro-Wilks test is a test performed to detect whether
contains 7 different independent variables which are: a variable is normally distributed or not depending on the p-
value. In case the p-value was less than or equal 0.05, the test
• Graduate Record Exam1 (GRE) score. The score will be out will reject the null hypothesis. Otherwise, the variable is
of 340 points. normally distributed. It is good to mention that Shapiro test does
has limitations. Moreover, it is biased toward large samples. The
• Test of English as a Foreigner Language2 (TOEFL) score, larger the sample, the more possibility to get a statistically
which will be out of 120 points. significant results [6].
• University Rating (Uni.Rating) that indicates the Bachelor B. Multiple Linear Regression
University ranking among the other universities. The score
will be out of 5 Multiple linear regression is a statistical technique used to
predict a dependent variable according to two or more

1
Graduate Record Examination www.ets.org/gre
2
Test Of English as Foreign Language www.ets.org/toefl
independent variables. As well as, present a linear relationship on the best model that showed the least MSE which was multi-
between them and fit them in a linear equation. The format of linear regression.
the linear equation is as following [7]:
In addition, Chakrabarty et al. [16] compared between both
yi=β0+β1xi1+...+βnxn +ϵ (1) linear regression and gradient boosting regression in predicting
chance of admit; point out that gradient boosting regression
where, for i=n observations: showed better results.
yi=dependent variable Gupta et al. [17] developed a model that studies the graduate
xi= independent variables admission process in American universities using machine
learning techniques. The purpose of this study was to guide
β0=y-intercept students in finding the best educational institution to apply for.
βn=slope coefficients for each independent variable Five machine learning models were built in this paper including
SVM (Linear Kernel), AdaBoost, and Logistic classifiers.
ϵ=the model’s error term or residuals
Waters and Miikkulainen [18] proposed a remarkable article
C. K-Nearest Neighbor that helps in ranking graduation admission application according
K-nearest neighbor (KNN) is a supervised machine learning to the level of acceptance and enhances the performance of
algorithm used for classification and regression problems [8], reviewing applications using statistical machine learning.
[9]. It is based on the theory of similarity measuring. Therefore,
Sujay [19] applied linear regression to predict the chance of
to predict a new value, neighbors should be put into
admitting graduate students in master’s programs as a
consideration. KNN uses some mathematical equations to
percentage. However, no more models were performed.
calculate the distance between points to find neighbors. In a
regression problem, KNN is used to find the mean of the k IV. DATASET PROCESSING AND FEATURE SELECTION
labels. While in classification problems, the mode of k labels
will be returned [10]. A. Correlated variables
D. Random Forest The dataset contained an independent variable to present the
serial number of the requests. According to my expertise, it does
The random forest algorithm is one of the most popular and not correlate to the dependent variable; hence, it was removed
powerful machine learning algorithms that is capable of from the dataset. It is good to mention that tests showed that
performing both regression and classification tasks [11]. This there are no missing values in any row of the database.
algorithm creates forests within number of decision reads.
Therefore, the more data is available the more accurate and B. Outliers
robust results will be provided [12]. Outliers are data values that differ greatly from the majority
Random Forest method can handle large datasets with higher of a set of data. To find the outliers, there are many methods that
dimensionality without overfitting the model. In addition, it can can be used, such as: scatterplot and boxplot. In this paper
handle the missing values and maintains accuracy of missing outliers will be investigated using boxplot method.
data [12]. As for the boxplot, the middle part of the plot represents the
E. Multilayer Perceptron first and third quartiles. The line near the middle of the box
represents the median. The whiskers on either side of the IQR
Multilayer perceptron is a supervised deep artificial neural represent the lowest and highest quartiles of the data. The ends
network machine-learning algorithm used to predict value of a of the whiskers represent the maximum and minimum of the
dependent variable of a dataset according to weights and bias data, and the individual circles beyond the whiskers represent
[13]. Weights are updated continuously when finding any error outliers in the dataset [20].
in classification. The first layer is the input layer, more than one
layer are presented next as hidden layers where each layer will
contain a linear relationship between the previous layer, and the
final layer is output layer that makes decisions and predicts [14].
Forward pass and backward pass can be performed. Forward
pass is where information flows from left to right, in other
words, the flow will be input, hidden layers, and output in order.
On the other hand, backward weights will adjust according to
the gradient flow in that direction [14].
III. RELATED WORK
A great number of researches and studies have been done on
graduation admission datasets using different types of machine
learning algorithms. One impressive work by Acharya et al. [15]
has compared between 4 different regression algorithms, which
are: Linear Regression, Support Vector Regression, Decision
Trees and Random Forest, to predict the chance of admit based Figure 1 Boxplot of Chance of Admit
 As noticed from Figure 1, there are outliers beneath the Note that the first and third quartile are found using the
boundaries of the boxplot. following equations where N is equal number of data points.
There are many different methods to deal with outliers, such
as: First Quartile (Q1)=(N+1) × 0.25 (2)

• Remove the case Third Quartile (Q3)=(N+1) × 0.75 (3)

• Assign the next value nearer to the median in place of
the outlier value V. MODEL DESIGN
• Calculate the mean of the remaining values without the A. Independent Variable Importance
outlier and assign that to the outlier case
To find the importance range of the independent variables,
In our case, there are few outliers; therefore, the best option Random Forest classifier can be used. The higher the value,
will be removing the rows that contain outliers. the more important it is.
C. Dataset Devision
Table 2 Independent Variable Importance
After cleaning the dataset, it was divided randomly into two
parts using holdout method. The first part contains 80% of the Independent Variable Rank
dataset to present training with 498 observations. The second
part contains 20% of the dataset to present testing with 98 GRE 1.6818745
observations.
TOEFL 1.3432065
D. Feature Selection
Uni. Rating 0.4589954
In order to perform feature selection, ols_step_best_subset()
function has been applied; which will display all possible SOP 0.7489156
subsets. Then according to certain criteria, the best subset will LOR 0.3707980
be selected. Since there are 7 independent variables, the number
of subsets to be tested is 27, which equals 128 subsets. To apply CGPA 2.6919350
feature selection, linear regression equation should be applied.
Note that linear regression can be performed only with numeric Research 0.2661699
independent variables.
The results in table 2 show that the most important variable
It is observed that all variables are numeric. It is good to
is CGPA as it has the highest ranking among all other variables.
mention that in case of having categorical variable, as.numeric()
And the second highest variable is GRE.
function can be used to convert data variables to numeric. Next,
regression model is created, and feature selection is performed. B. Histogram
The best model in feature selection is presented by the model A histogram plot is used to present the frequencies of
with either highest R-square or smallest MSEP, which belongs continuous numbers and to show the distribution of the data
to model 6 which includes all columns except column 4 that selected. Outliers and skewness can be predicted from a
presents SOP. histogram along with some other features. Skewness measures
how much a graph is asymmetric.
E. Descriptive Summary
Descriptive summary provides numerical measures of some
important features which describe a dataset. Some features are
presented in table 1:
Table 1 Descriptive Summary

Feature Value
Minimum Value 0.3600
First Quartile (Q1) 0.6325
Median 0.7200
Mean 0.7233
Third Quartile (Q3) 0.8200
Maximum Value 0.9700
Figure 2 Histogram of Chance of Admit
 Figure 2 shows the histogram graph of the dependent Table 3 displays the p-value of each parameter obtained from
variable Chance of Admit with skewness −0.25 to the left. Shapiro-Wilk test; all three parameters’ p-value are less than
0.05. Therefore, the null hypothesis is rejected, and variables are
The histogram graphs of the most important independent not normally distributed.
variables are presented also according to the importance test.
D. Multicollinearity Issue
Multicollinearity is a huge issue that exists whenever an
independent variable is highly correlated with one or more
independent variables in a multiple regression equation. If VIF
is > 10, high multicollinearity is found. This problem can lead to
unstable regression model. In other words, any slight change in
the data will lead to a huge change in the coefficients of the
multiple linear regression model [21], [22].
In conclusion, there is no multicollinearity problem since all
the values are less than 10. This also leads to the fact that our
regression model is stable.
E. Linear Regression
According to the linear regression model applied, the
equation that represents regression model is:
Figure 3 Histogram of CGPA Regression model= -1.33 + )0.002 * GRE + 0.0026 * TOEFL +
0.005* Uni.Rating + 0.004 * SOP +0.013 * LOR + 0.118 *
Figure 3 shows the histogram of CGPA, the most important CGPA + 0.023 * Research)
independent variable, with skewness −0.0283553 to the left.
According to Pr(>|t|) value from the linear regression test,
all variables have a statistically significant role except for
columns 3, 4, which are Uni.Rating and SOP. Also, the R-
squared value = 0.83. which means that 83% of variation in our
dataset can be explained with our model. The p-value is 2.2e-16,
which is way less than 0.05 so we reject the null hypothesis and
the model is statistically significant.
VI. RESULTS AND DISCUSSIONS
A. Statistical Test
According to the normality test, the dependent variable is not
normally distributed. Therefore, nonparametric test will be
performed using PHStat. The test is one-way ANOVA which is
performed to determine whether three samples or more have any
statistically significant differences between their means or not
[23].
Figure 4 Histogram of GRE
The test shows that p-value equals 0.97, which is greater than
Figure 4 shows the histogram of GRE with skewness −0.04 0.05, thus, the null hypothesis cannot be rejected, and the tests
to the left. are not statistically different.
C. Normality Test B. Mean absolute error
Normality test shows whether a parameter is normally The different regression models are performed on
distributed or not. Shapiro test is used to perform normality test. Admission dataset through Weka in order to decide which model
If the p-value is greater than 0.05, this means it is normally performs the best based on mean absolute error (MAE) value.
distributed. Otherwise, the graph is not normally distributed. The results are shown in Table 4:
Table 3 Shapiro-Wilk Normality Test Table 4 Performance Analysis

Parameter p-value Regression model MAE value

Chance of Admit 3.239e-6 Multi linear regression 0.0343
CGPA 0.01171 Random Forest 0.0363
GRE 0.0001245 KNN 0.0544
Multilayer perceptron 0.0337
 According to table 4, multilayer perceptron has the smallest “Machine Learning Classifications of Coronary Artery Disease,” in
MAE equivalent to 3.37% which means that it is the best model. 2018 International Joint Symposium on Artificial Intelligence and
Natural Language Processing (iSAI-NLP), 2018, pp. 1–6.
CONCLUSION
[10] N. S. Altman, “An introduction to kernel and nearest-neighbor
In this paper, machine learning models were performed to
nonparametric regression,” Am. Stat., vol. 46, no. 3, pp. 175–185,
predict the opportunity of a student to get admitted to a master’s
1992.
program. The machine learning models included are multiple
linear regression, k-nearest neighbor, random forest, and [11] A. B. Nassif, M. Azzeh, L. F. Capretz, and D. Ho, “A comparison
Multilayer Perceptron. Experiments show that the Multilayer between decision trees and decision tree forest models for software
Perceptron model surpasses other models. development effort estimation,” in 2013 3rd International
Conference on Communications and Information Technology,
As for the future work, more models can be conducted on
more datasets to learn the model that gives the best performance. ICCIT 2013, 2013, pp. 220–224.
[12] T. K. Ho, Random Decision Forests. USA: IEEE Computer Society,
REFERENCES 1995.
[1] M. Injadat, A. Moubayed, A. B. Nassif, and A. Shami, “Multi-split [13] A. B. Nassif, “Software Size and Effort Estimation from Use Case
Optimized Bagging Ensemble Model Selection for Multi-class Diagrams Using Regression and Soft Computing Models,”
Educational Data Mining,” Appl. Intell., vol. 50, pp. 4506–4528, University of Western Ontario, 2012.
2020. [14] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning
[2] F. Salo, M. Injadat, A. Moubayed, A. B. Nassif, and A. Essex, internal representations by error propagation,” MIT Press.
“Clustering Enabled Classification using Ensemble Feature Cambridge, MA, vol. 1, no. V, pp. 318–362, 1986.
Selection for Intrusion Detection,” in 2019 International [15] M. S. Acharya, A. Armaan, and A. S. Antony, “A comparison of
Conference on Computing, Networking and Communications regression models for prediction of graduate admissions,” ICCIDS
(ICNC), 2019, pp. 276–281. 2019 - 2nd Int. Conf. Comput. Intell. Data Sci. Proc., pp. 1–5, 2019.
[3] M. N. Injadat, A. Moubayed, A. B. Nassif, and A. Shami, [16] N. Chakrabarty, S. Chowdhury, and S. Rana, “A Statistical
“Systematic ensemble model selection approach for educational Approach to Graduate Admissions’ Chance Prediction,” no. March,
data mining,” Knowledge-Based Syst., vol. 200, p. 105992, Jul. pp. 145–154, 2020.
2020.
[17] N. Gupta, A. Sawhney, and D. Roth, “Will i Get in? Modeling the
[4] A. Moubayed, M. Injadat, A. B. Nassif, H. Lutfiyya, and A. Shami, Graduate Admission Process for American Universities,” IEEE Int.
“E-Learning: Challenges and Research Opportunities Using Conf. Data Min. Work. ICDMW, vol. 0, pp. 631–638, 2016.
Machine Learning Data Analytics,” IEEE Access, 2018.
[18] A. Waters and R. Miikkulainen, “GRADE : Graduate Admissions,”
[5] M. S. Acharya, A. Armaan, and A. S. Antony, “A Comparison of pp. 64–75, 2014.
Regression Models for Prediction of Graduate Admissions,”
[19] S. Sujay, “Supervised Machine Learning Modelling & Analysis for
Kaggle, 2018. .
Graduate Admission Prediction,” vol. 7, no. 4, pp. 5–7, 2020.
[6] S. S. Shapiro, M. B. Wilk, and B. T. Laboratories, “An analysis of
[20] G. Singler, “Statistics Reference Series Part 1: Descriptive
variance test for normality,” 1965.
Statistics,” 2018.
[7] G. K. Uyanık and N. Güler, “A Study on Multiple Linear
[21] D. E. Farrar and R. R. Glauber, “Multicollinearity in regression
Regression Analysis,” Procedia - Soc. Behav. Sci., vol. 106, pp.
analysis; the problem revisited,” no. 1, pp. 5–7, 2003.
234–240, 2013.
[22] R. M. O’Brien, “A caution regarding rules of thumb for variance
[8] C. López-Martín, Y. Villuendas-Rey, M. Azzeh, A. Bou Nassif, and
inflation factors,” Qual. Quant., vol. 41, no. 5, 2007.
S. Banitaan, “Transformed k-nearest neighborhood output distance
[23] E. Ostertagová and O. Ostertag, “Methodology and Application of
minimization for predicting the defect density of software projects,”
Oneway ANOVA,” Am. J. Mech. Eng., vol. 1, no. 7, pp. 256–261,
J. Syst. Softw., vol. 167, p. 110592, Sep. 2020.
2013.
[9] A. B. Nassif, O. Mahdi, Q. Nasir, M. A. Talib, and M. Azzeh,

Authors Contribution: Sara Aljasmi wrote the paper. Ali Bou Nassif, Ismail Shahin and Ashraf Elnagar
revised the experiments and the whole paper

View publication stats