ARTIFICIAL NEURAL NETWORK (ANN) MODEL FOR PREDICTING STUDENTS ACADEMIC PERFORMANCE

Journal of Science and Information Technology October 2013, Vol. 1 No.2, PP. 23-37
1
1
Usman, O.L., 2
Adenubi, A.O.
1,2
Department of Computer Science, Tai Solarin University of Education, Ogun State Nigeria.
+234 (0) 807 391 1635, +234 (0) 805 868 4616
(usmanol@tasued.edu.ng, adenubiao@tasued.edu.ng)
USMAN O.L: He is an academic staff of the Computer Science Department of the Tai
Solarin University of Education, Ogun State, Nigeria. He holds a Master and Bachelor’s Degrees
in Computer Science Education from the Tai Solarin University of Education, Nigeria.
ADENUBI, A.O: He is a member of the academic staff in Computer Science Department of
the Tai Solarin University of Education, Ogun State, Nigeria. He holds a Master Degree in
Management of Information Technology from the University of Nottingham, United Kingdom
with a Bachelor Degree in Computer Science from the Olabisi Onabanjo University, Nigeria.

2
ARTIFICIAL NEURAL NETWORK (ANN) MODEL FOR PREDICTING
STUDENTS’ ACADEMIC PERFORMANCE
Abstract
The observed poor quality of graduates of some Nigerian Universities in recent times is traceable
to non-availability of mechanism that would enable the University administrators to project into
the future performance of the concerned students. This will guarantee the provision of better
educational services as well as customize assistance according to students’ predicted level of
performance. In this research, Artificial Neural Networks (ANNs) were used to develop a model
for predicting the final grade of a university student before graduating such student. The data
used in this study consists of thirty (30) randomly selected students in the Department of
Computer Science, Tai Solarin University of Education in Ogun State, who have completed four
academic sessions from the university. Test data evaluation showed that the ANN model is able
to correctly predict the final grade of students with 92.7% accuracy. All ANN models used were
trained and simulated using nntool of MATLAB (2008a) software.
Index: Neural Network, Artificial Intelligence, Student Achievement Prediction, Student,
Academic Performance

3
ARTIFICIAL NEURAL NETWORK (ANN) MODEL FOR PREDICTING
STUDENTS’ ACADEMIC PERFORMANCE
1.0 INTRODUCTION
Advising students on their class performance and motivating them in order to improve on their
performance is an integral part of every instruction. The mechanisms to achieve the above aim
required a technique capable of accurately predicting student achievement as early as possible
and cluster them for better academic assistance. According to Lykourentzou et al, (2009),
student-achievement prediction can help identify the weak learners and properly assist them to
cope with their academic pursuit. Several methods and systems have been developed for the
above task, most of which are artificial intelligence-based. For instance, Lykourentzou et al.,
(2009) estimated the final grades of students in e-learning courses with multiple feed-forward
neural networks using multiple-choice test data of students of National Technical University of
Athens, Greece as input. The results obtained shows that ANN is 91.2% efficient. Junemann,
Lagos, and Arriagada (2007) used neural networks to predict future student schooling
performance based on students’ family, social, and wealth characteristics. The aforementioned
work focused on predicting the achievement of 15-year-old secondary students on reading,
mathematics and science subjects in Berlin.
In the Nigeria context, Oladokun, Adebanjo & Charles-Owaba (2008) applied multilayer
perceptron neural network for predicting the likely performance of candidates being considered
for admission into Engineering Course of the University of Ibadan using various influencing
factors such as ordinary level subjects’ scores, matriculation exam scores, age on admission,

4
parental background etc., as input variables. The results showed that ANN model is able to
correctly predict the performance of more than 70% of prospective students.
However, Abass et al., (2011) applied another technique of Artificial Intelligence (AI) i.e., case-
base reasoning (CBR) to predict student academic performance based on the previous datasets
using 20 students in the Department of Computer Science, TASUED as the study domain. The
high correlation coefficient observed between the actual graduating CGPA and the CBR
predicted ones also justify the usefulness and effectiveness of AI techniques in this type of task.
In this research work, Artificial Neural Network is used to estimate students’ final grade in the
university with a prediction level of 92%.
2.0 ARTIFICIAL NEURAL NETWORK
Inspired by the structure of the brain, an Artificial Neural Network (ANN) consists of a set of
highly interconnected entities, called Processing Elements (PE) or unit. Each unit is designed to
mimic its biological counterpart, the neuron. Each accepts a weighted set of inputs and responds
with an output. Neural Networks address problem that are often difficult for traditional
computers to solve, such as speech and pattern recognition, weather forecasts, sales forecasts,
scheduling of buses, power loading forecasts and early cancer detection. The origin of the neural
network can be traced to 1940s when two researchers, Warren McCulloch and Walter Pitts, tried
to build a model to simulate how biological neurons work. Though the focus of this research was
on the anatomy of the brain, it turns out that this model introduced a new approach for solving
technical problem outside neurobiology. Neural networks have been applied in clustering,
pattern recognition, function approximation and prediction systems. Several architectures for the

5
ANN exist. These include feed-forward, feed-backward, single-layer, recurrent, radial basis
function network, and self-organizing maps.
Among the neural network architectures, feed-forward is most commonly used. Feed- forward
neural networks (FFNN) tend to be straight forward network that associate inputs with outputs.
According to Haykin (1999), FFNN consists of one or more hidden layers of neurons. In this
type of network, neuron connections, called synapses, do not form a directed cycle. The goal of
the FFNN training is to minimize a cost function typically defined as the mean square error
(MSE) between its actual and target outputs, by adjusting the network synaptic weights and
neuron parameters. More specifically, these network parameters are adjusted based on the back-
propagation algorithm. In this algorithm, information is passed forward from the input nodes,
through the hidden layers, to the output nodes, and the error between the expected and the
network response is calculated. Then, this error signal is propagated backwards to the input
neurons. A popular approach to optimize the performance of back-propagation is the Levenberg-
Marquardt algorithm, which has been found to increase the speed convergence and effectiveness
of the network training (Hagan & Menhaj, 1994; Lykourentzou et al., 2009). Typical example is
TRAINLM. Other important approach is Gradient Descent algorithm, for example, TRAINGDM
and TRAINGDA. By this approach, input vectors are applied to the network and calculated
gradients at each training sample are added to determine the change in synaptic weights and
biases (Haykin, 1999; Folorunso et al., 2010). The FFNN parameters are estimated based only
on the training dataset, and the performance of the network is evaluated by computing the MSE
on the validation dataset. They are extensively used in pattern association, function
approximation, prediction and data clustering.

6
3.0 METHODOLOGY FOR PREDICTING STUDENTS’ ACADEMIC PERFORMANCE
WITH ANN
The methodology for predicting students’ performance and designing a tool for performing this
task is clearly divided into seven (7) recognizable steps as captured in the flowchart below
(Fig.1) and the resulting experimental data is presented in Table 1. In this study, TRAINLM,
TRAINGDM and TRAINGDA functions were used to train FFNN. The objective is to determine
which of the training functions will produce the best results. TRAINLM function updates weight
and bias values of FFNN according to Levenberg-Marquardt optimization, whereas
TRAINGDM and TRAINGDA functions update FFNN weight and bias values according to the
Gradient Descent optimization. While TARINGDM is a gradient descent with momentum, the
TARINGDA is a gradient descent with adaptive learning rate. The subject in this study consists
of 30 randomly selected students that have already completed four academic sessions with the
University. The samples are in the age range (22-25 years), cut-across all intelligent levels, and
are exposed to the same learning experience of the case study. The entire dataset are divided into
two sets:
The Input Variable
The input variables are the dataset used as input to the ANN models constructed in the study, as
well as the target values used to compare the predicted values against the reality. The first three
(3) sessions CGPA values of the samples were used as inputs while the final CGPA values
served as target values.

7
The Output Variable
The output variable represents the performance of a student on graduation. The output variable is
based on the current grading system of the case study. The classifications of output variable
domain are: 1st
Class as “Distinction”, 2nd
Class (Upper Division) as “Very Good”, 2nd
Class
(Lower Division) as “Good”, 3rd
Class as “Fair”, and Pass/Fail as “Poor”.
Fig.1: Flowchart representation of Methods for predicting students’ academic performance

8
STUDENT
ID
TARGET
FINAL
CGPA Interpretation
ANN1
OUTPUT Interpretation
ANN2
ANN3
TRAINLM TRAINGDM TRAINGDA
Student1 2.38 Fair 1.83 Fair 2.47 Good 2.39 Fair
student2 4.20 Very Good 4.20 Very Good 4.02 Very Good 4.08 Very Good
Student3 3.08 Good 4.20 Very Good 3.09 Good 3.06 Good
Student4 3.39 Good 4.20 Very Good 3.57 Very Good 3.49 Good
Student6 4.12 Very Good 4.20 Very Good 3.98 Very Good 4.02 Very Good
Student7 2.11 Fair 1.83 Fair 2.31 Fair 2.14 Fair
Student8 2.88 Good 1.83 Fair 2.60 Good 2.88 Good
Student16 2.31 Fair 2.33 Fair 2.39 Fair 2.58 Good
Student21 2.14 Fair 1.84 Fair 2.66 Good 2.89 Good
Student23 2.50 Good 2.92 Good 2.46 Good 2.55 Good
Student24 3.57 Very Good 4.20 Very Good 3.44 Good 3.29 Good
Student25 2.86 Good 4.20 Very Good 3.42 Good 3.55 Very Good
Student26 3.36 Good 4.20 Very Good 3.64 Very Good 3.78 Very Good
Table 1: Experimental Results

9
3.1 Analysis of ANN model Trained with TRAINLM Function (ANN1)
From Table 1, it can be observed that no student was given either Distinction or Fail grade by
ANN1. According to this training function, 14 students were given Very Good grades as against
6 students who graduated with Very Good. Also, only 1 student was predicted with Good as
against 14 students who graduated with Good. The number of students predicted with Fair final
grade is 15 whereas only 10 students graduated with Fair. Out of the 30 students’ grades that
were used in the experiment only 17 students grades were predicted correctly, though their
values are seem to be uncorrelated while 13 students’ grades were wrongly predicated. The
percentage accuracy of this training function is 56.7%. The performance plot (Fig.2) of this
training function shows that the training function is not efficient at minimizing the mean square
error (MSE) between its responses and the actual students’ final grade. Fig.3 depicts the
relationship between the response of ANN1 trained with this function (vertical axis) and the
desire response (horizontal axis). The regression plot in Fig.3 showed that student final grade
prediction is possible at a correlation coefficient R value equal to 0.89687 though the training
time is 7sec. This function completed the training cycle in 136 iterations.
Fig.2: ANN1 Training (TRAINLM) Fig.3: Testing ANN1

10
3.2 Analysis of ANN model Trained with TRAINGDM Function (ANN2)
Clearly from Table 1, no value was predicted for Distinction and Fail which is also true
according to the actual final grade. Very Good grade was predicted for 7 students as final grades
whereas only 6 students graduated with Very Good. ANN2 predicted Good grade for 17 students
as their final grade while 14 students actually graduated with Good. Also, the model predicted
that only 6 students would graduate with Fair as their final grade whereas 10 students graduated
with Fair. This training function showed considerable improvement as the model generated from
it correctly predicted the final grade of 23 students out 30 students used in the experiment
bringing its percentage accuracy as high as 76.7%. From the graph in Fig.4 and 5, one can
deduce that the training function is efficient at minimizing the performance criterion (mean
square error) between its responses and the students’ final grades. The regression plot in Fig.5
showed that student final grade prediction is possible at a correlation coefficient R value equal to
0.98331 though the training time was long i.e., 17sec and the Epoch is 1000 iterations. It can be
observed here that this training function completed its training in very long time using the
maximum number of training iterations.
Fig. 4: ANN2 Training (TRAINGDM) Fig.5: Testing ANN2

11
3.3 Analysis of ANN model Trained with TRAINGDA Function (ANN3)
The results from the training of a neural network (ANN3) trained with TRAINGDA function as
presented in Table 1. According to this table, there was no prediction for Distinction and Fail
grade in line with the actual performance of students under consideration. Very Good grade was
predicted for 7 students as final grades, like TRAINGDM function whereas only 6 students
graduated with Very Good in their final year. ANN3 predicted Good grade for 15 students as
their final grade while 14 students actually graduated with Good. We observe from the last
statement that, this is almost correct as the difference between the students who graduated with
Good and predicted number is 1. Also, the model predicted that only 8 instead of 10 would
graduate with Fair grades in their final year. From above analysis, one can conclude that this
training function is efficient than those mentioned in the previous sections and the model
(ANN3) generated from it is very efficient at minimizing the performance criterion (mean square
error) between its responses and the students’ final grades as shown by performance graph of
Fig.6. With this training function, ANN3 was able to correctly predict the final grade of 25
students out 30 students used in the experiment giving the percentage accuracy of 83.3%. The
regression plot in Fig.7 showed that student final grade prediction is possible at a correlation
coefficient R value equal to 0.98813. This function is very efficient in computational time (2 sec)
and training epoch (105 iterations) with validation check of 6. The output predicted by this
model is use in the design of our ANNSPP. The summary of these analyses is shown in Table 2.

12
Fig. 6: ANN2 Training (TRAINGDA) Fig.7: Testing ANN3
ANN1 ANN2 ANN3
Summary
Original
Grade
Predicted
Grade
Original
Grade
Predicted
Grade
Original
Grade
Predicted
Grade
Distinction 0 0 0 0 0 0
Very
Good 6 14 6 7 6 7
Good 14 1 14 17 14 15
Fair 10 15 10 6 10 8
Fail 0 0 0 0 0 0
Total 30 30 30 30 30 30
Correct 17 23 25
Incorrect 13 7 5
%
Accuracy 56.70% 76.70% 83.30%
Table 2: Summary of ANNs Performance
The best trained network was then used to design an interface called Artificial Neural Network
for Students’ Performance Prediction (ANNSPP) using Visual Basic platform of Visual Studio 8.
The designed interface is implemented to predict the likely final grades of some set of students
when supplied with unknown CGPA values.

13
4.0 DISCUSSION OF RESULTS
From the various tests performed on the results of the training, validation and test results, it is
confirmed that Artificial Neural Network (ANN) performs quite impressible in estimating the
Final Grades of students in university. Both the percentage accuracies and correlation
coefficients are good evidences of the fact that, given appropriate data at its disposal, the
ANNSPP designed can ensure students’ learning outcome prediction accuracy and help the
stakeholders in education sector and university management to dynamic.
lly group students according to their predicted level of performance and offer better educational
services to them. The study also corroborates earlier researches that have reported the
effectiveness of ANN in predicting learners’ achievements at various levels and forms of
education.
5.0 CONCLUSION AND FUTURE WORK
The result obtained from the study actually showed that the Artificial Neural Networks are
capable to predict the performance of students in the university and can be used to develop a
predictive tool. This is due to a little but not all that significance errors which exists between the
training values and the ANN simulated values. Suffice to say that the positive and negative 0.3
experienced in the study can be reduced if the number of times the Cumulative Grade Point
Aggregate (CGPA) values is used in the training session increases. From the results, it can be
concluded that ANN is 92.7% efficient at predicting student final academic performance.

14
REFERENCES
Abass,O., Oyekanlu ,E.A., Alaba, O.B,and Longe. O.B.,(2011) ”Forecasting Student Academic
Performance using Case-Base Reasoning”. International conference on ICT for Africa.
23-26th
march, 2010 Otta Nigeria, pp.105-112.
Folorunso, O., Akinwale, A.T., Asiribo, O.E., and Adeyemo, T.A.(2010). Population Prediction
Using Artificial Neural Network. Africa Journal of Mathematics and Computer Science
Research, Vol.3(8).pp.155-162.
Hagan, M.T., & Menhaj, M.B. (1994). Training feed-forward networks with the Marquardt
algorithm. IEEE Transactions on Neural Networks, 5, 989–993.
Haykin, S.(1999),Neural Networks: A Comprehensive Foundation (2nd
Edition). Macmillan
College Publishing Company, New York.
Junemann, M.A.P., Lagos, P.A.S., & Arriagada, R.C. (2007). Neural networks to predict
schooling failure/success. In J. Mira & J.R. Álvarez (Eds.), Lecture Notes in Computer
Science, Vol. 4528: Nature Inspired Problem-Solving Methods in Knowledge
Engineering (pp. 571–579). Berlin: Springer.
Karamouzis, S. (2000). Using Analogies to predict Student performance. In Proc.
IASTED’s Artificial Intelligence and soft comp. Banf, Canada, pp.355-360.
Lykourentzou, I., Giannoukos, I., Mpardis, G., Nikolopoulos, V., and Loumos, V. (2009),
Early and Dynamic Student Achievement Prediction in E-Learning Courses Using Neural
Networks. Journal of the American Society for Information Science and Technology, Vol.
60(2):372-380.
McCulloch, W.S., and Pitts, W.A., (1943). A logical Calculus of Ideas Imminent in Nervous
Activitity. The Bulletin of Mathematical Biophysics, Vol.5, pp.115-133. Doi:
10.1007/BF02478259.
Oladokun, V.O Adebanjo, A.T. and Charles-Owaba, O.E. (2008). Predicting Students Academic
Performance using Artificial Neural Network: A case study of an Engineering Course.
Pacific Journal of Science and Technology, Vol.9(1):72-79.
Zhang, G., Patuwo, B.E (1998), Forecasting with Artificial Neural Networks: The State of the
Art. International Journal of Forecasting, Vol.14, pp.35-62.

ARTIFICIAL NEURAL NETWORK (ANN) MODEL FOR PREDICTING STUDENTS ACADEMIC PERFORMANCE

More Related Content

Similar to ARTIFICIAL NEURAL NETWORK (ANN) MODEL FOR PREDICTING STUDENTS ACADEMIC PERFORMANCE (20)

More from Emily Smith (20)

Recently uploaded (20)

ARTIFICIAL NEURAL NETWORK (ANN) MODEL FOR PREDICTING STUDENTS ACADEMIC PERFORMANCE