A Model for Predicting Students’ Academic Performance using a Hybrid of K-means and Decision tree Algorithms

International Journal of Computer Applications Technology and Research
Volume 4– Issue 9, 693 - 697, 2015, ISSN: 2319–8656
www.ijcat.com 693
A Model for Predicting Students’ Academic Performance
using a Hybrid of K-means and Decision tree Algorithms
Thaddeus Matundura Ogwoka
School of Open, Distance and
eLearning (SoDeL)
Jomo Kenyatta University of
Agriculture and Technology
(JKUAT)
Nairobi, Kenya
Wilson Cheruiyot
School of Computer Science
and Information Technology
(JKUAT)
Nairobi, Kenya
George Okeyo
School of Computer Science
and Information Technology
(JKUAT)
Nairobi, Kenya
Abstract: Higher learning institutions nowadays operate in a more complex and competitive due to a high demand from prospective
students and an emerging increase of universities both public and private. Management of Universities face challenges and concerns of
predicting students’ academic performance in to put mechanisms in place prior enough for their improvement. This research aims at
employing Decision tree and K-means data mining algorithms to model an approach to predict the performance of students in advance
so as to devise mechanisms of alleviating student dropout rates and improve on performance. In Kenya for example, there has been
witnessed an increase student enrolling in universities since the Government started free primary education. Therefore the Government
expects an increased workforce of professionals from these institutions without compromising quality so as to achieve its millennium
development and vision 2030. Backlog of students not finishing their studies in stipulated time due to poor performance is another
issue that can be addressed from the results of this research since predicting student performance in advance will enable University
management to devise ways of assisting weak students and even make more decisions on how to select students for particular courses.
Previous studies have been done Educational Data Mining mostly focusing on factors affecting students’ performance and also used
different algorithms in predicting students’ performance. In all these researches, accuracy of prediction is key and what researchers
look forward to try and improve.
Keywords: Data Mining; Decision tree; K-means; Educational Data Mining
1. INTRODUCTION
To predict how students may perform during their learning
process is a complex task despite continuous increase of data
in the databases relating to students academics in institutions
of higher learning. According to (Marquez et al., 2013), the
academic management systems are not designed properly to
support educational managers to investigate which students
are at risk of dropping out of university. Data Mining is the
process of discovering interesting patterns and knowledge
from large amounts of data (Han and Kamber, 2003). From
educational Data Mining (EDM) website , “Educational Data
Mining is an emerging discipline , concerned with developing
methods for exploring the unique types of data that come from
educational settings, and use using them to better understand
students and the setting in which they learn” . To analyze
students’ learning process is complex, but thanks to EDM in
which its methods and approaches can be used to predict
students’ performance like the model being proposed in this
research paper. Hence university managers will have options
of to come with strategies of improving student academic
performance (Borka and Rajeswari, 2013). This research
proposes the use of Decision tree classification and K-means
clustering algorithms to develop a model for predicting the
academic performance of students in higher level institutions
like universities. Prediction is a method of carrying out
Educational Data Mining (EDM) using clustering algorithms
like K-means and classification algorithms like decision trees
to predict student performance (Ramesh et al., 2013). In
university, the performance a student at the end of every
semester determines whether a student is to progress to the
nest academic year which leads to the completion of his/her
studies. Passing in these semester examinations is crucial
since it will determine whether a student is to get to final year
which later realizes a student graduating and ushered into the
economic development of a country. This is one of the
reasons institutions of higher learning are established for
(Patel et al., 2013). In this research paper, WEKA knowledge
analysis software tool is used to for the analysis the
algorithms used and the model performance.
2. RELATED WORK
Great work has been done and is always being done by this
area of Educational Data Mining. From (Shovon and Haque,
2012)’s research, where they used k-means algorithm to
predict the student learning activities by clustering them into:
“Good”, “Medium”, and “Low” based on their GPA. They
used 50 students as the training samples and concluded that
their prediction accuracy was low and needed improvement in
future. Our proposed model has realized improved accuracy
by using 173 students as the algorithm training samples.
(Yedav and Pal, 2012) using decision trees’ ID3, CART, and
C4.5 classifiers, conducted a study to predict student
academic performance and realized an accuracy of 62.22%,
62.27%, and 67.77% respectively. In our research, using
decision tree’s J48, we realized a prediction accuracy of
98.8439% on the student training instances. (Kalpesh and Pal,
2013) in their model of predicting students’ performance
using decision tree’s ID3 and C4.5 on 173 training datasets,
achieved a prediction accuracy of 75.145% at 47.6
milliseconds execution time.
2.1 Data Clustering
Data clustering is unsupervised statistical analysis technique,
which is used to segment large data into homogeneous groups
called clusters, in order to discover hidden patterns and
relationships to help in quick decision making (Shovon and

www.ijcat.com 694
Haque, 2012). K-means is simple algorithm that partitions “n”
observations into k clusters in which each member belongs to
a cluster of nearest mean (Mustafa et al., 2010).
Figure 1. Basic K-means algorithm (Mustafa et al., 2010)
2.2 Decision tree classification algorithm
Decision tree is a data mining technique which can be applied
in prediction tasks, it is a tree-like structure in which the root
and each internal node are labelled with a question (Shovon
and Haque, 2012). It is a classifier with the following
structure: Decision node specifies a test on a single attribute,
Leaf node indicates the value of the target attribute, Arc/edge
split of an attribute, and Path is a junction test to make final
decision.
3. PROPOSED MODEL APPROACH
In The proposed model is to take into account the educational
data mining requirements as cited by other authors (Baker and
Yasef, 2009) of keeping continuous record of student
progress. The goal of this study is to predict students’
academic performance. The approach will contain among
others: Data tier which is the collection of student semester
records, Application layer that will contain data extraction,
loading, and knowledge repository. Then finally the
presentation layer for displaying results analyzed to the user.
3.1 Data processing and analysis
In data mining, before running tests on the collected data
instances, it is necessary to clean and prepare the data for use.
In our research, sample data from the Technical University of
Mombasa-Kenya student management system was cleaned to
look at the relevance of the data attributes to be able to
remove any redundancy, or irrelevant features, and analyzed
using WEKA software tool. In our study, we have considered
dataset of undergraduate students pursuing Bachelor of
Science in Technology (BTIT), Bachelor science in
Information Technology (BSIT) both government sponsored
and self-sponsored, part-time and full-time students at the
department of computer science and information technology
(CSIT) in the Technical university of Mombasa which was
our case study. According to (Suchita and Rajeswari, 2013),
on the basis of the data collected some variable attributes are
considered to predict student academic performance:
Attendance%, Assignment%, Unit tests% and University
result%. But for the purpose of this research, only useful
fields from the single combined table were selected for our
study. Some of the selected variables from the database and
recommended attributes for academic prediction are shown
below.
Table 1. Selected variables from student records
3.2 Analysis of data
Here we are evaluating the Decision tree algorithm using J48
classifier and K-means algorithm using Simple-Kmeans
option from WEKA. From our framework architecture, it
consisted of Data Tier where data from students is collected
and interesting variables selected. Application tier where data
cleaning and mining using algorithms is done producing a
prediction model. We used First semester results to train the
algorithm (this was the training Dataset), the second semester
final results was predicted by the algorithms (this was our
Test dataset). After successful learning, the algorithms were
tested for prediction of second semester results in which the
final grade column was left blank for the algorithm to predict
the students’ performance (predicted results).
Figure 2. Part of test dataset used for prediction
Part of the training data used for training the algorithms (first
semester student records-cleaned) is shown below:
Figure 3. Part of training data set
This are part of the student datasets which is used to train the
algorithms in our model. It represents the first semester
student results. The test dataset of records represents second
semester students’ records in which the model is supposed to
predict how the student will perform. And use the model to

www.ijcat.com 695
cluster students in groups according to their predicted result
relationships.
4. RESULTS & DISCUSIONS
Here we evaluated the experiments conducted using the
algorithms (Decision tree using J48 and K-means) the analysis
and results was done in WEKA. The decision tree algorithm
was used to do the prediction and show in tree-like structure
the results which will help in making decisions from the
predictions made. The K-means algorithms run on the same
dataset on the same WEKA tool, was used to group the
predicted students into several groups of our choice in relation
to the attributes of the records as shown below. Several
classification metrics are used to evaluate the results
(Manahaes et al., 2013):
 Accuracy : the measure of correctly classified
instances
 True positive (TP): the proportion of positive
cases (P) correctly classified as such and
 True Negative (TN): the proportion of negative
cases (F) correctly classified as such.
After loading the training dataset into WEKA, the
preprocessing yielded the following results as shown in Figure
4 below.
Figure 4. preprocessed training datasset results
From preprocessing results of training dataset, 143 students
are identified as passed-P (in blue color as per the bar graph)
and 30 students as failed-F (in red color as per the bar graph)
in the fig 1above. This is true as per the original dataset, we
had a total of 173 student samples. After preprocessing, by
using Decision tree’s J48 classifier, the algorithm was trained
using the same dataset. The results are as shown in fig 5
below.
From the training result, 171 intances were classified correctly
with an accuracy of 98.8439%, while 2 instances were
incorectly classified with an error of 1.1561%. The kappa
statistic which takes account of similarities between classsses
was 0.95978% which is better. :A kappa value greater than
zero indicates the classifier is doing better than chance”
(Manhaes et al, 2011). TP Rate: Rate of true positives (these
indicated students who were correctly clasified in relation to
final grade as the class)-the model realised 0.988 out of 1. FP
Rate: Rate of false positives (these indicated students who
were wrongly classified as belonging to a given final grade as
aclass)-the model generated 0.029 out of 1. Precision: these
represents a proportion of students that are truly of
aclass(given final grade) divided by the total students
classified as that class-the model generated 0.988 out of 1.
Recall: these represents the proportion of students from the
experiment who were classified as a given class (final grade)
diveided by the actual total in that class (eqivalent to TP
Rate)-the model generated 0.988 out of 1. F-measure: this is
acombined measure of Predcision and Recall calculated as
2*Precision*Recall/(Precision+Recall) and it was 0.988. From
all these acuracy measures, they all approached 1, hence
showing our algorithm was learning well from the training
instances.
The figure 6 below shows decision tree view generated by the
model
Figure 6 decision tree view of the model

www.ijcat.com 696
The view above shows in “tree”-structure one of the views
that can be used in decision making. From the above results,
30 students whose mean marks were <=49 failed (F) while
those with >49 mean marks in their first semester results
passed (P).
4.1 Prediction results
After learning using first semester students results, we tested
the algorithm with same students but on different results
(second semester records) and the Class attribute
”FinalGrade” was left blank for the algorithm to predict as
was shown in figure 2. Using decision tree algorithm, on the
test data, the following output were realized as predicted
results as shown in figure 7 below.
Figure 7 part of predicted results from the model
As shown from figure 7 above, the model was able to predict
the second semester’s final grades (as shown by arrow), and
as shown from the decision tree view of the predicted results
in figure 8 below, 30 students who failed (F) had their end of
semester two mean mark <=49 while 143 students passed with
a mean mark of >49.
Figure 8 predicted results decision tree view
4.2 K-means clustering predicted results
From the same model, using WEKA, the following results
were output.
Figure 9 predicted results clusters
“Clustering is the process of grouping a set of elements in
such a way that the elements in the same group or cluster are
more similar to each other than to those in other groups or
cluster” (Kalpesh et al., 2013). As shown from the above fig
4.8, a student of admission number “BTIT/0061/2013M2”
who is a male, entry mode by JAB, learning full-time,
belonging to cluster0 passed with 53 mean mark, his
attendance was good and had sat for the continuous
assessment test.
4.3 Results summary
According to (Kalpesh et al., 2013) who predicted students’
academic performance using ID3 and C4.5 algorithms,
obtained the following accuracy as compared to what we
obtained using J48 and K-Means algorithms as shown in table
2 and table 3 below.
Table 2 Algorithm accuracy. Source: (Ogwoka et al., 2015)
Algorithm Total
studen
ts
Correctly
predicted
students
Accurac
y (%)
Execution
time in
millisecon
ds
J45 173 171 98.8439 20
Table 3 Algorithm accuracy. Source: (Kalpesh et al.; 2013)
Algorithm Total
studen
ts
Correctly
predicted
students
Accurac
y (%)
Execution
time in
millisecon
ds
ID3 173 130 75.145 47.6
C4.5 130 75.145 39.1

www.ijcat.com 697
5. SUMMARY OF FINDINGS
The main objective of our research was to apply decision tree
and k-means algorithms to create a model for predicting
students’ performance. To achieve this, our model was
realized through stepwise specific objectives:
1. We evaluated decision tree and k-means algorithms
in terms of their operations using WEKA free
software tool and other written literature.
2. We successfully applied decision tree and k-means
algorithms and created a model of predicting
students’ academic performance where we analyzed
173 undergraduate students of Technical University
of Mombasa’s Computing and information
technology department using first semester results
to predict second semester results.
3. We tested the model of predicting students’
academic performance and realized an accuracy of
98.8439% at an execution time of 20 milliseconds.
From the most previous researchers we looked into, their
biggest challenge was to have an increased accuracy. All the
their models as was seen in the previous literature, their
accuracies were less than 90% with an execution time as big
as 47.6 milliseconds (Kalpesh et al., 2013). Our results of 0.05
milliseconds execution time and accuracy of 98.8439% has
reduced this gap.
6. CONCLUSIONS
A model for predicting students’ academic performance using
Decision tree and k-means algorithms has an improved
accuracy and easily be implemented in institutions of higher
to do prediction of students’ performance and also mine
interesting features pertaining academics of students.
7. RECOMMENDATIONS AND
FUTURE WORK
From the results and findings of the experiments done in this
study, the researcher recommends the adoption of student
performance prediction models as Education Data Mining is
an emerging data Mining discipline. In our research, WEKA
does not update automatically on test dataset predicted as is
the case on training dataset, hence to view the results you
have to save in a file. In future, we will explore if WEKA has
improved on this feature to use in our model or research more
on more other open source data mining and analysis tools on
this recommendation.
8. REFERENCES
[1] Marquec-vera, C. Cano, A. Romero, C., and Ventura, S.
Predicting student failure at school using genetic
programming and different data mining approaches with
high dimensional and imbalanced data. Applied
Intelligence, 2013, 1:1-16.
[2] Han, J., Kamber, M., and Pei, J. (2013). Data Mining
Concepts and Techniques, third edition. Morgan
Kaufmann.
[3] JEDM-Journal of Educational Data Mining, (JEDM.
ISSN2157-2100). Meaning of Educational Data Mining.
Retrieved from
www.educationaldatamining.org/JEDM/index.php/JED
M.
[4] Burka, S., and Rajeswari, K. Predicting Student
Academic Performance using Data Mining, 2013, 2(7),
213-219.
[5] Ramesh, P., Parkavi, P., and Ramar, K Predicting
Student Performance. A statistical and Data Mining
approach, 2013, 63(8), 0975-8887.
[6] Patel, M., Abdul, K., and Pappal, P. Educational Data
Mining and its Role in Education Field, 2013, 5(2),
2458-2461.
[7] Ajay, P., and Saurabh, P., Data Mining Techniques in
EDM for Predicting the Performance of Students, 2013,
2(6), 2279-0764.
[8] Shovon, M., and Haque, M., An approach of Improving
Student Academic Performance by using K-means
clustering Algorithm and Decision tree, 2012, 3:8
[9] Ahmed, A., and Elaraby, I., Data Mining: A prediction of
Student performance using classification method, 2014,
2:43-47.
[10] Mustafa, T., Ayesha, S., and Khan, M. 2010. Data
Mining Model for Higher Education System. European
Journal of Scientific, 1450-216X.

A Model for Predicting Students’ Academic Performance using a Hybrid of K-means and Decision tree Algorithms

More Related Content

What's hot (19)

Viewers also liked (13)

Similar to A Model for Predicting Students’ Academic Performance using a Hybrid of K-means and Decision tree Algorithms (20)

More from Editor IJCATR (20)

Recently uploaded (20)

A Model for Predicting Students’ Academic Performance using a Hybrid of K-means and Decision tree Algorithms