(IJACSA) International Journal of Advanced Computer Science and Applications,

Vol. 11, No. 9, 2020

A Predictive Model for the Determination of

Academic Performance in Private Higher Education
Institutions
Francis Makombe1, Manoj Lall2
Department of Computer Science
Tshwane University of Technology
Pretoria, South Africa

Abstract—The growth and development of predictive models goal of reducing dropouts, increasing retention, performance
in the current world has influenced considerable changes. Today, and graduation rate.
predictive modelling of academic performance has transformed
more than a few institutions by improving their students' Application of the appropriate data mining technique that
academic performance. This paper presents a computational suits the current scenario is important in order to identify
predictive model using artificial neural networks to predict useful patterns. In this article, factors that have an impact on
whether a student will pass or fail. The model is unique in the the pass rates of students are identified and used in the
current literature as it is specifically designed to evaluate the classification model. The following algorithms are applied in
effectiveness of the predictive strategies on neural networks as the construction of the classification model-Artificial Neural
well as on five additional algorithms. The analysis of the Networks, Logic Regression, eXtremeGBoost, SVM, Naive
experimental results shows that Artificial Neural Networks Bayes, and Random Forest algorithms.
outperformed the eXtremeGBoost, Linear Regression, Support
Vector Machine, Naive Bayes, and Random Forest algorithms for The rest of this article is structured as follows: the
academic performance prediction. literature review is presented in Section II. The description of
the data and the methodology used are presented in
Keywords—Classification modelling; data mining; higher Sections III and IV. The results and its discussion are
education institutions; accuracy; academic performance presented in Section V. In Section VI, conclusions and
recommendations are presented.
I. INTRODUCTION
Public higher education providers are institutions that have II. LITERATURE REVIEW
been established and funded by the state through the In a research conducted by [4], the researchers attempted
Department of Higher Education and Training (DHET). Public to explore the applicability of Fuzzy C-Means clustering
providers include universities, universities of technology, and technique for academic performance of students. They found
comprehensive universities. Private providers are owned by that fuzzy C-Means clustering algorithm serves as a good
private organizations or individuals. Higher education benchmark to monitor the progression of students modelling
institutions (HEIs) operate in an increasingly complex and in educational domain. The author in [5] also recommended a
challenging environment. Competition has increased, and fuzzy logic-based expert system that periodically evaluates
previously anticipated government funding has become scarce student performance and supplies students with feedback on
[1]. In such circumstances, HEIs must succeed in a financial progress within data grid environment. The system made use
sense or else they will go out of business [2]. In their quest for of the fuzzy logic theory and develop the decision making
survival, common practices adopted by HEIs are to increase process based on fuzzy rules to assess whether a student gets
the intake of students and try to improve on their success rates. very poor, poor, good, average or excellent performance.
Since, many government and private funds depends on the
throughput rates of institutions, being able to predict the In an attempt to identify the main attributes that may affect
chances of any new student’s success is very important. This the performance of students in engineering, [6] applied data
study aims to improve the pass rates of students’ in a mining concepts such as k-Means clustering and Decision tree
particular private academic institution by providing a Techniques. They used records of 1500 students enrolled for
classification model to assist in identifying student at risk of various subjects in engineering. The author in [7] investigated
failing a program. Being able to identify such students, the the impact of classroom attendance and gender on academic
educational institutes can provide a targeted support performance of university students in an Organic Chemistry
mechanisms to the needy students. The author in [3] mention course. Data was collected through survey involving real time
that the reasons for the identification of a student at risk of documentation of attendance for each student at each class
dropouts or attrition early enough are to be able to provide lesson over a three month period. Their findings show that
necessary support and interventions for the student with the attendance had a significant impact on the performance. In
another study, [8] analysed the impact of class attendance,
practical work and assignments in a course on the success rate.

415 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 11, No. 9, 2020

They found that the number of given assignment has a  Number of employed parents or guardians.
negative impact on the academic performance. They used
C4.5 as the classification algorithm for their work. Several  Group Assignment marks.
other studies conducted have shown that class attendance is an  Test marks.
important predictor of academic outcomes which conclude
that students who attend more classes generally earn higher  Individual Assignment marks.
final grades [9].
The scatterplot (Fig. 1) shows the distribution of individual
In a study by [10], one of the factors that influences a test marks in relation to the individual assignment marks. In
student’s ability to succeed is the socioeconomic conditions. analysis of this scatterplot, most of the students perform well
This fact is supported by [11] who state that Student poverty in both tests and individual assignments. There are a few
and the lack of sufficient funding have consistently been cited outliers who perform very well in individual assignments but
as key reasons for student academic failure and progression poorly in tests. According to this scatterplot, the approximate
difficulties. In the study by [12], they used marks of four range for tests with most students’ marks is 40 to 80, and that
academic batches of Computer Science & Information for the individual assignments is 50 to 90. This shows that
Technology (CS&IT) students for predicting performance. In students are generally performing better in individual
their study, they collected records of 347 undergraduate assignments than in tests.
students have been mined with classifiers such as Decision
The scatterplot (Fig. 2) for Test and Group assignment
tree, Neural Networks and Naive Bayes.
marks shows that a greater proportion of students perform
In another study, [13] applied Naïve Bayes for the very well in group assignments, where they take part in
classification of student evaluation. Their dataset consisted of research activities. By comparison, a lot of students fail the
the following parameter-age, place of birth, gender, high tests as shown by the large concentration of test marks below
school status (public or private), department in high school, the mark of 50, compared to the test mark greater than 50.
organization activeness, age at the start of high school level, This could provide a basis for intervention by the private
and progress GPA score. institution in efforts to assist the students prepare better for
tests.
Discriminate analysis was done by [14] to predict the
success and failure of students in a specific physics course.
Discriminate analysis is a similar technique to multiple
regression except that it is used for categorized data. They
used this technique to provide a function that contains the
variables that should be used for predicting the success of a
student. They collected the data for 1622 students who
enrolled in Electricity & Magnetism course, which had a high
rate of failure. At first they identified many possible predictors
such as, SAT grade, MATH GPA, Overall GPA. In another
study [15], applied predictive modelling techniques to identity
students at risk of dropping out of their registered
qualification. They used Support Vector Machine, Naïve
Bayes, Decision tree, K-nearest neighbors and Random Forest
on 1156 students.
III. DATA DESCRIPTION
Fig. 1. Scatterplot of Test and Individual Assignment marks.
This research followed a quantitative approach.
Questionnaires were administered to private academic
institutions in an anonymously manner to enhance the privacy
and anonymity of the participants. The questionnaires in this
study were distributed in two ways: manually handed out and
also using the online survey tool survey monkey. The dataset
consisted of the following attributes:
 Study hours per week.
 Bursary - whether a student has a bursary or not.
 Class Attendance.
 Student workload (number of modules registered).
 Fulltime study or attending through part-time classes.
 English language proficiency marks. Fig. 2. Scatterplot of Test and Group Assignment Marks.

416 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 11, No. 9, 2020

IV. METHODOLOGY E. Naïve Bayes Classifiers

In order to assess the effect of data quality, data attribute These refer to a collection of "probabilistic classifiers"
significance and class number in the academic performance which are based on the application of Bayes ‘theorem with
prediction in this study, six classification algorithms have strict (Naïve) independence assumptions amongst features.
been selected and implemented in R programming language. Naïve Bayes classifiers are very scalable.
These algorithms were chosen because they cover the different
F. eXtreme Gradient Boosting
approaches used by classifiers for learning and they are state
of the art algorithms that are often used in data mining eXtreme Gradient Boosting (XGBoost) is a versatile and
applications [16]. enhanced gradient algorithm booster variant designed for
efficiency, machine speed and performance of the model. It is
A. Random Forests an ensemble learning technique that combines multiple
Instead of building a single tree for classification, the machine learning algorithms to lessen errors and increase
Random Forests constructs a set of trees, and uses them all to prediction accuracies.
classify or to predict. Random Forests where developed by
[17], and they create a (forest) collection of decision trees by V. RESULTS AND DISCUSSION
the method of bagging. Random Forests are sets of learning The following chart demonstrates the different accuracy,
models where the unknown input is listed according to the sensitivity, and F-measure values obtained (Fig. 3).
majority vote of decision-making bodies. This means that the Inaccuracies are also shown for each of the six algorithms
class predicted by most of the trees would be the last class in used in this research. Fig. 1 shows that neural networks
the set. Random Forests, increase the classification algorithm had the best accuracies which also had the least
performance, avoiding overfitting and are robust to outliers inaccuracies. It also had high precision, and F-measure values
and noise [17]. where a good classifier has an F-measure value of close to 1,
whilst the worst classifier has an F-measure close to 0.
B. Neural Network- Multilayer Perceptron (MLP)
This refers to an artificial neural feed network class in A. Receiver Operating Characteristic (ROC) Curve
which at least three layers of nodes are present: one input The purpose of the Receiver Operating Characteristic
layer, one hidden layer and one output layer. Every node is a (ROC) curve is to primarily assess the accuracy of a
non-linear activation neuron except for the input nodes. MLP continuous measurement that is performing a binary outcome
uses a supervised learning method called training prediction. The best classifier has an area under the curve
backpropagation [18]. Every node layer is fully connected to (AUC) value close to 1. Fig. 4 shown below are the AUC
the next layer, which generates a finite acyclic graph values for three classifiers, two with the best performance and
(DAG).Except the input node, each node is a processing node one with a poor performance.
that is used to calculate the output based on an input using a
non-linear activation function. Each link of two nodes has a The following values where obtained for the AUC. This
change in weight depending on the training data set. The was done for three classifiers, which where the two best
weight adjustments are based on the error of the measured classifiers, and the worst classifier Table I.
output difference and the predicted output. The weights are
adjusted to reduce the error by using a gradient descent.
C. Support Vector Machines (SVMs)
SVMs for binary classification where developed by [19].
This is an approach that is used to solve classification
problems using linear methods for both datasets having
linearly and nonlinearly separable classes [20].
D. Linear Regression
Linear regression helps to predict the value of the Y
outcomes variable on the basis of one or more X variables
(Equation 1). The objective is to create a linear relationship (a
mathematical formula) between a predictor variable(s) as well
as the response variable, such that the value of the Y answer is
determined by using this formula only when the values (Xs) of
the predictors are known. In general, the formula for linear
regression is provided as follows:
Y = β1 + β2X + ϵ (1)
Fig. 3. Performance Statistics for different Algorithms used.
where, β1 is the intercept and β2 is the slope. These are
called regression coefficients, and ϵ is the error term, which
refers to the area of Y, that the regression model cannot be
able to explain.

417 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 11, No. 9, 2020

Fig. 5 shows the predictions of a sample of six students

using the neural networks. These are the computed values
which show the predicted value of whether a student will pass
or fail a module. The simple learning rate algorithm was used
for these predictions. The value of 0.4566725 for the first
student in the dataset means that the student is more likely to
fail this module. Similarly, the value of 0.6010540 (which is
greater than 0.5) for the second student would mean that this
student is more likely to pass this module.

Fig. 5. Output of Neural Networks.

VI. CONCLUSIONS AND RECOMMENDATIONS

Fig. 4. ROC for Two Most Accurate and the Least Accurate Classifier.
In the study the researcher shows the degree of accuracy of
the six algorithms used in the study, and their related
TABLE I. AUC VALUES FOR TWO MOST ACCURATE CLASSIFIERS AND
THE LEAST ACCURATE CLASSIFIER misclassification errors. It was observed that ANN performed
better than Logic Regression, eXtremeGBoost, SVM, Naive
Algorithm AUC Bayes, and Random Forest algorithms. It was observed that
eXtremeGBoost 0.62 bursary and group assignments had a positive correlation with
the pass rate. The recommendations, based on the results
Neural networks 0.86
obtained, are: (1) Group assignments have a positive
SVM 0.43 correlation concerning whether a student will pass or fail as
they have a direct effect. Hence it is recommended that
By making use of the AUC and accuracy values obtained students should be encouraged to take a more active role in
in this experiment the neural networks algorithm was selected group assignments. (2) Bursaries have a positive correlation
to be the most suitable algorithm for the prediction of with academic performance; therefore, it is recommended for
academic performance for this study. The performance of the private institute to provide bursaries to successful
ANN was followed by eXtremeGBoost and then SVM. It can applicants. (3) There should be provision made for booster or
therefore be concluded that the neural net algorithm support classes meant for students predicted to fail. To have a
outperformed the other five algorithms for academic more accurate assessment of a student’s academic
performance classification. performance, data from other domains of higher education
B. Confusion Matrix Results value chain such as psychosocial domain, cognitive domain,
institutional domain, personality domain, and demographic
Table II below summarizes the experimental results domain should be considered as future work.
obtained for both the training and testing dataset, and it also
demonstrates the accuracies and misclassification errors REFERENCES
obtained using a neural network defined with the simple [1] E. J. Dumond and T. W. Johnson, "Managing university business
learning rate algorithm. educational quality: ISO or AACSB?," Quality Assurance in Education,
2013.
TABLE II. NEURAL NETWORK ALGORITHM WITH SIMPLE LEARNING [2] H. J. Juhl and M. Christensen, "Quality management in a Danish
RATE CONFUSION MATRIX RESULTS business school–A head of department perspective," Total Quality
Management, vol. 19, no. 7-8, pp. 719-732, 2008.
CONFUSION MATRIX [3] O. W. Adejo and T. Connolly, "Predicting student academic
performance using multi-model heterogeneous ensemble approach,"
True False False True Misclassification Journal of Applied Research in Higher Education, 2018.
Dataset
Positive Positive Negative Negative Error
[4] R. S. Yadav and V. P. Singh, "Modeling academic performance
Training evaluation using fuzzy c-means clustering techniques," International
5478 88 86 3636 0.019
data Journal of Computer Applications, vol. 60, no. 8, 2012.
Test [5] S. Patel, P. Sajja, and A. Patel, "Fuzzy logic based expert system for
1223 17 23 1059 0.017 students performance evaluation in data grid environment," International
data
Journal of Scientific & Engineering Research, vol. 5, no. 1, 2014.

418 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 11, No. 9, 2020

[6] V. Sreenivasarao and C. G. Yohannes, "Improving academic [13] N. Dengen, E. Budiman, M. Wati, and U. Hairah, "Student Academic
performance of students of defence university based on data Evaluation using Naïve Bayes Classifier Algorithm," in 2018 2nd East
warehousing and data mining," Global Journal of computer science and Indonesia Conference on Computer and Information Technology
technology, 2012. (EIConCIT), 2018: IEEE, pp. 104-107.
[7] O. D. Ayodele, "Class attendance and academic performance of second [14] E. W. Thomas, M. J. Marr, A. Thomas, R. M. Hume, and N. Walker,
year university students in an organic chemistry course," African Journal "Using discriminant analysis to identify students at risk," in Technology-
of Chemical Education, vol. 7, no. 1, pp. 63-75, 2017. Based Re-Engineering Engineering Education Proceedings of Frontiers
[8] N. A. Yassein, R. G. M. Helali, and S. B. Mohomad, "Predicting student in Education FIE'96 26th Annual Conference, 1996, vol. 1: IEEE, pp.
academic performance in KSA using data mining techniques," Journal 185-188.
of Information Technology and Software Engineering, vol. 7, no. 5, pp. [15] R. Lottering, R. Hans, and M. Lall, "A model for the identification of
1-5, 2017. students at risk of dropout at a university of technology," in 2020
[9] A. Kirby and B. McElroy, "The effect of attendance on grade for first International Conference on Artificial Intelligence, Big Data, Computing
year economics students in University College Cork," Vol. XX, No. XX, and Data Communication Systems (icABCD), 2020: IEEE, pp. 1-8.
Issue, Year, 2003. [16] I. H. Witten and E. Frank, "Data mining: practical machine learning
[10] D. E. Roby, "Research on school attendance and student achievement: A tools and techniques with Java implementations," Acm Sigmod Record,
study of Ohio schools," Educational Research Quarterly, vol. 28, no. 1, vol. 31, no. 1, pp. 76-77, 2002.
pp. 3-16, 2004. [17] L. Breiman, "Random forests," Machine learning, vol. 45, no. 1, pp. 5-
[11] S. Mngomezulu, R. Dhunpath, and N. Munro, "Does financial assistance 32, 2001.
undermine academic success? Experiences of'at risk'students in a South [18] J. Gao, X. He, and L. Deng, "Deep learning for web search and natural
African university," Journal of Education (University of KwaZulu- language processing," 2015.
Natal), no. 68, pp. 131-148, 2017. [19] C. Cortes and V. Vapnik, "Support-vector networks," Machine learning,
[12] R. Asif, A. Merceron, and M. K. Pathan, "Predicting student academic vol. 20, no. 3, pp. 273-297, 1995.
performance at degree level: a case study," International Journal of [20] D.-M. Tsai and C.-C. Lin, "Fuzzy C-means based clustering for linearly
Intelligent Systems and Applications, vol. 7, no. 1, p. 49, 2014. and nonlinearly separable data," Pattern recognition, vol. 44, no. 8, pp.
1750-1760, 2011.

419 | P a g e
www.ijacsa.thesai.org