Performance Evaluation of Feature Selection Algorithms in Educational Data Mining

Integrated Intelligent Research (IIR) International Journal of Data Mining Techniques and Applications
Volume: 05 Issue: 02 December 2016 Page No.131-139
ISSN: 2278-2419
131
Performance Evaluation of Feature Selection
Algorithms in Educational Data Mining
T.Velmurugan1
, C. Anuradha2
1
Associate Professor, PG and Research Dept. of Computer Science, D.G.Vaishnav College, Chennai, India.
2
Research Scholar, BharathiarUniversity, Coimbatore, India.
Email:
velmurugan_dgvc@yahoo.co.in;anumphil14@gmail.com;
Abstract: Educational Data mining(EDM)is a prominent field
concerned with developing methods for exploring the unique
and increasingly large scale data that come from educational
settings and using those methods to better understand students
in which they learn. It has been proved in various studies and
by the previous study by the authors that data mining
techniques find widespread applications in the educational
decision making process for improving the performance of
students in higher educational institutions. Classification
techniques assumes significant importance in the machine
learning tasks and are mostly employed in the prediction
related problems. In machine learning problems, feature
selection techniques are used to reduce the attributes of the
class variables by removing the redundant and irrelevant
features from the dataset. The aim of this research work is to
compares the performance of various feature selection
techniques is done using WEKA tool in the prediction of
students’ performance in the final semester examination using
different classification algorithms. Particularly J48, Naïve
Bayes, Bayes Net, IBk, OneR, and JRip are used in this
research work. The dataset for the study were collected from
the student’s performance report of a private college in Tamil
Nadu state of India. The effectiveness of various feature
selection algorithms was compared with six classifiers and the
results are discussed. The results of this study shows that the
accuracy of IBK is 99.680% which is found to be high than
other classifiers over the CFS subset evaluator. Also found that
overall accuracy of CFS subset evaluator seems to be high than
other feature selection algorithms. The future work will
concentrate on the implementation of a proposed hybrid
method by considering large dataset collected from many
institutions.
Keywords: Educational data mining, Wrapper selection, Best
First Search,Classification Algorithms, Feature selection
Algorithms.
I. INTRODUCTION
In educational data mining, prediction of students’
performance has long been an interesting area of research and
it helps to identify the weak students or students at risk. As the
educational institutions are facing intense competition with
respect to admission, retention and sustenance, it is important
that the institutions pay significant attention in improving
students output. Most often the institutions are judged by the
percentage of results produced by students’ in the finale end
semester examination. A data mining system offer several
techniques to the educational leaders to support the decision
making process to improve the quality of education. The large
volume of data generated in the educational institutions can be
effectively used to draw rich source of vital information to
support decision making system. The main focus of this
research work is to identify the best feature selection and
classification algorithms to examine a performance of
undergraduate student performance in education data set. The
objective is to find the best attribute by comparing the
performance of various feature selection techniques in the
prediction of students’ performance in the final semester
examination using different classification algorithms such as
J48, Naïve Bayes, Bayes Net, IBk, OneR, and JRip are used in
this research work. The idea behind this research work is to
identify slow learners which help the faculties to give special
attention to individual student’s to improve their academic
performance.
A research work done by parneet kaur et al. in education
sector. Their work focuses on identifying the slow learners
among students and applies feature selection algorithms to
filter desired potential variables using WEKA tool. As a result,
statistics are generated based on all classification algorithms in
order to predict the accuracy [*]. Another work by
HythemHashim et al. discussed about Data mining
methodologies to study student’s academic performance using
the C4.5 Algorithm. Their objective is to build a classification
model that can be used to improve the student’s academic
records in Faculty of Mathematical Science and Statistics. This
model has been done using C4.5 for predicting student
performance in many different settings [1].
A work done by Vaibhav and Rajendra named as Classification
and performance evaluation using data mining algorithms. The
authors collected student data from polytechnique institute and
classified the data using Decision tree and Naïve Bayesian
algorithms. They compare results of classification with respect
to different performance parameters [2]. Another research done
by Anjana and Jeena discussed about Predicting College
Students Dropout using EDM Techniques. Here WEKA tool
has been used to evaluate the attributes. Various classification
techniques like induction rules and decision tree have been
applied to data and results of each of these approaches have
been compared [3]. A paper Titled “Performance
Analysis and Prediction in Educational Data Mining: A
Research Travelogue” by Pooja et al. has been done towards
the usage of data mining techniques in the field of education.
This paper presents a comprehensive survey towards
educational data mining [4]. A work by Punlumjeak and
Rachburee had proposed a comparison of feature selection
techniques namely genetic algorithms, support vector machine,
information gain, minimum and maximum relevance
algorithms with supervised classifiers such as naïve bayes,
decision tree, k-nearest neighbour and neural network. Their
results shows that minimum and maximum relevance feature
selection method with 10 features give the best result on

ISSN: 2278-2419
132
91.12% accuracy with k-nearest neighbour classifier[5].
Another work by Anal and Devadatta had applied a different
feature selection algorithm on the student data set. The best
results are achieved by correlation based feature selection with
8 features. Subsequently classification algorithms may be
applied on this feature subset for predicting student grades
[6].Komal and Supriya [7] have conducted a Survey on Mining
Educational Data to Forecast Failure of Engineering Students.
This paper provides a Review of the available literature on
Educational Data mining, Classification method and different
feature selection techniques that author should apply on
Student dataset. The research paper titled Improvement on
Classification Models of Multiple Classes through Effectual
Processes by Tarik [8].
This paper work focuses on improving the results of
classification models of multiple classes via some effective
techniques. The collected data are pre-processed, cleaned,
filtered, normalized, the final data was balanced and
randomized, then a combining technique of Naïve Base
Classifier and Best First Search algorithms are used to
ultimately reduce the number of features in data sets. Finally, a
multi-classification task is conducted through some effective
classifiers such as K-Nearest Neighbor, Radial Basis Function,
and Artificial Neural Network to forecast the students’
performance. Another work carried out by Sadaf and Kulkarni
discussed about Precognition of Students Academic Failure
Using Data Mining Techniques. This research paper proposes
to pre-recognize Student’s academic failure using various Data
mining techniques especially induction rules, decision trees
and naïve Bayes are applied [9].
Carlos et al. [10] have tried to attempt to solve this problem of
predicting student’s academic failure using clustering
algorithms, induction rules or decision trees algorithms of data
mining techniques. Authors applied five rules of induction
rules and five decision tree algorithms on the dataset. Sahil and
Shweta have carried out a Study of Application of Data Mining
and Analytics in Education Domain. This paper basically is a
study of certain research experiments which aims to study the
different applications of data mining techniques on the
educational data. Also it elaborated upon the state of the art
techniques in the field of educational analytics [11]. Ogunde
and Ajibade have developed a new system for the prediction of
students graduation grades based on entry results data. The
proposed system uses ID3 algorithm to classify the data and
construct the decision tree by employing a top-down, greedy
search to test every attributes [12].Dinesh and Radika had done
a survey on predicting Student academic Performance in
educational environment which is based upon the
psychological and environmental factor is predicted using
different educational data mining techniques. Researchers also
survey the predictive model in data mining and current trends
in prediction in data mining [13].
A Work done by ArpitTrivedi has put forward a simple
approach for categorizing student data using decision tree
based approach. For taking measures of category of specific
student, a frequency measure is used as a feature extraction.
With the use of trained classifier, they predicted the class for
indefinite student automatically [14].A work has done by
Agrawal and Gurav have done a review on Data Mining
Techniques Used for Educational System. This paper is based
on survey which proposes to apply data mining techniques
such as association rule mining, classification techniques
[15].The classification is a data mining technique which
includes systematic approach to building the classification
models from an input dataset [16]. Some of the popular
classifiers used to solve a classification problem are decision
tree classifiers, rule-based classifiers, neural networks, support
vector machines, and naive Bayes classifiers [17]. Therefore, a
key objective of the learning algorithm is to build a predictive
model that accurately predicts the class labels of previously
unknown records. This paper examines that various
classification algorithms and their performance are compared
using WEKA software and results are discussed. The open
source data mining tool WEKA was used in the present work
to obtain the reduced set of features from the available feature
set using various feature selection techniques. In addition, the
reduced attributes were given as input to the classifiers like
decision tree algorithm C4.5 (J48), Bayesian classifiers like
Naïve Bayes and BayesNet, Nearest Neighbor algorithm (IBk)
and rule learners (OneR and JRip) to evaluate the performance
of the classification algorithms for the particular feature
selection technique.
This paper is structured as follows. Section 2 discusses about
background of the study. Section 3 describes various feature
selection techniques used for reducing the attributes of the
dataset. The statement of the problem is provided in Section 4.
The details of the dataset generated for the study is presented
in the Section 5. The experimental evaluation and comparative
analysis are given in Section 6 and Conclusion for the
proposed work is given in Section 7. Finally, vital references
are mentioned in Section 8.
II. BACKGROUND
Feature selection has been an important field of research in
data mining and machine learning systems. The primary
objective of any feature selection technique is to choose a
subset of features of the input variables by eliminating those
features which are redundant, irrelevant orof no predictive
information [18]. Feature subset selection in machine learning
can be broadly classified into three groups as filter, wrapper
and embedded models [19]. Filters based method of feature
selection depends on the general characteristics of the training
data. Thus, feature selection process is carried out as a pre-
processing step, independent of the learning algorithm.
Wrapper technique depends on the learning algorithm and uses
it as a black box to evaluate the usefulness of subsets of
variables in the prediction task. Thus, wrapper methods uses
learning algorithm to evaluate the subset of features for feature
selection. Wrapper methods are computationally intensive.
Embedded methods on the other hand perform feature selection
during the training process of the classifier. This methods are
particularly specific to a given learning machines.
As the dimensionality of a domain expands, the number of
features N increases. Finding an optimal feature subset is
intractable and problems related feature selections have been
proved to be NP-hard. At this juncture, it is essential to
describe traditional feature selection process, which consists of
four basic steps, namely, subset generation, subset evaluation,
stopping criterion, and validation. Subset generation is a search
process that produces candidate feature subsets for evaluation

ISSN: 2278-2419
133
based on a certain search strategy. Each candidate subset is
evaluated and compared with the previous best one according
to a certain evaluation. If the new subset turns to be better, it
replaces best one. This process is repeated until a given
stopping condition is satisfied [20]. A number of studies have
established in theory and practice that feature selection is an
effective technique in improving learning efficiency,
enhancing predictive accuracy and minimizing the complexity
of results in data mining system. The effectiveness of feature
selection has been proved in many applications involving data
mining and machine learning like text categorization [21],
image retrieval [22], information retrieval [23], DNA
microarray analysis [24], intrusion detection [25,26], and
music information retrieval [27].
III. STATEMENT OF THE PROBLEM
In this research work, the performance of various feature
selection algorithms was evaluated on different classification
algorithm using the students’ academic performance dataset
generated for the study. The proposed study made several
comparisons to evaluate the effectiveness of the feature
selection techniques using the measures involving error and
accuracy parameters. The overall aim of the study was to
analyze the effectiveness of various machine learning
algorithms to predict students’ performance in the end semester
examination. The dataset for the study included the
demographic details of the students like gender, family size
and type, income, parent’s educational attainment and locality.
In addition, pre-collegiate conditions of the students like their
performance in secondary and higher secondary classes are
also collected and maintained in the colleges. Thus, it could be
useful to the educational leaders and management of the
colleges, if the features in the currently available data can be
acting as the indicator for predicting the performance of the
students. The major objective of this study is to analyze the
student’s data available in the degree colleges to identify any
specific patterns that might be useful in the prediction of their
performance in the university exams. The specific objective of
the study is to classify students according to their performance
in the final examination based on their personal and pre-
collegiate characteristics.
IV. RESEARCH METHODOLOGY
In this research work six classification algorithm are used such
as J48, Naïve bayes, Bayes net, IBK, OneR and JRip along
with four feature selection algorithms. In this section, the
fundamentals of some the feature selection algorithms are
illustrated. Furthermore, the algorithms CfsSubset evaluations,
Chi-Squared Attribute Evaluation, Information Gain Attribute
Evaluation and Relief attribute evaluation which are used in
this research work are also described.
A. Correlation-based Feature Selection (CFS)
Correlation based Feature Selection (CFS) is a simple filter
algorithm that ranks feature subsets according to a correlation
based heuristic evaluation function [28]. In CFS, the bias of the
evaluation function is toward subsets that contain features that
are highly correlated with the output to be predicted and
uncorrelated with each other. Irrelevant features should be
ignored because they will have low correlation with the class.
Redundant features should be screened out as they will be
highly correlated with one or more of the remaining features.
The acceptance of a feature will depend on the extent to which
it predicts classes in areas of the instance space not already
predicted by other features.
B. Best First Search Algorithm (BFS)
Best First Search is an approach that searches the attribute
subsets space via a method of Greedy Hill Climbing improved
with a backtracking aptitude. The controls of the amount of
backtracking can be achieved via setting the quantity of
consecutive non-improving nodes. This approach might start to
search in both directions; forwardly or backwardly. It can start
with the empty set of attributes and search forwardly, or it can
start with the full set of attributes and search backwardly
[8].The Table 1 shows the Best first search algorithm [28].
Table 1: Best first search algorithm
1. Begin with the OPEN list containing the start state,
the CLOSED list empty, and BEST← start state.
2. Let s = arg max e(x) (get the state from OPEN with
the highest evaluation).
3. Remove s from OPEN and add to CLOSED.
4. If e(s) _ e (BEST), then BEST ← s.
5. For each child t of s that is not in the OPEN
or CLOSED list, evaluate and add to OPEN.
6. If BEST changed in the last set of expansions, goto 2.
7. Return BEST.
C. Wrapper Feature Selection
In the wrapper approach, the feature subset selection is done
using the induction algorithm as a black box. The feature
subset selection algorithm conducts a search for a good subset
using the induction algorithm itself as part of the evaluation
function. The accuracy of the induced classifiers is estimated
using accuracy estimation techniques. Wrappers are based on
hypothesis. They assign some values to weight vectors, and
compare the performance of a learning algorithm with different
weight vector. In wrapper method, the weights of features are
determined by how well the specific feature settings perform in
classification learning. The algorithm iteratively adjusts feature
weights based on its performance [29].
D. CfsSubset Evaluator (CSER)
It evaluates the worth of a subset of attributes by considering
the individual predictive ability of each feature along with the
degree of redundancy between them [7].
E. Chi-Squared Attribute Evaluator (CSAER)
ChiSquaredAttributeEval evaluates an attribute by computing
the value of the chi-squared statistic with respect to the class
[7].
F. Information Gain Attribute Evaluator (IGAER)
It Evaluates an attribute by measuring the information gain
with respect to the class Info Gain (Class, Attribute) = H
(Class) - H (Class | Attribute) [7].
G. Relief Attribute Evaluator (RAER)
It Evaluates the worth of an attribute by repeatedly sampling an
instance and considering the value of the given attribute for the
nearest instance of the same and different class can operate on
both discrete and continuous class data [7].

ISSN: 2278-2419
134
Table 2: Description of options and Capability of CfsSubset
evaluator
Option Description
locally Predictive
Identify locally predictive attributes.
Iteratively adds attributes with the
highest mutual relationship with the
class as long as there is not already
an attribute in the subset that has a
higher correlation with the attribute
in question
missingSeperate Take missing as a separate value.
Capability Supported
Class Missing class values, Numeric class,
nominal class, Date class, Binary
class
Attributes Empty nominal attributed, Nominal
attributes, Numeric Attributes,
Unary attributes, Date attributed,
Binary attributes, Missing Values
Min # of instances 1
Table 3: Description of options and Capability of Chi-Squared
Attribute Evaluator
Option Description
BinarizeNumericAttributes Only binarize numeric
attributes instead of properly
discretizing them.
Missing Merge Distribute the counts for
missing values. Then counts
are distributed across other
values in proportion to their
frequency. Or else, missing is
treated as a separate value.
Class Missing class values, nominal
class, Binary class
Attributes Empty nominal attributed,
Nominal attributes, Numeric
Attributes, Unary attributes,
Date attributed, Binary
attributes, Missing Values
Table 4: Description of options and Capability of Info Gain
Attribute Evaluation
Option Description
binarizeNumericAttributes
Just binarize numeric
attributes rather than properly
discretizing them
Missing Merge
Distribute the counts for
missing values. Counts are
distributed over other
values in proportion to
their frequency. Or else,
missing is treated as a
separate value.
Class Missing class values, nominal
class, Binary class
Attributes Empty nominal attributed,
Nominal attributes, Numeric
Attributes, Unary attributes,
Date attributed, Binary
attributes, Missing Values
Table 5: Description of options and Capability of Relief
Attribute Evaluation
Option Description
numNeighbors Number of nearest neighbors for
attribute estimation
sample Size Number of instances to sample.
Default (-1) indicates that all
instances will be used for
attribute estimation.
Seed Random seed for sampling
instances
Sigma Set influence of nearest
neighbors. Used in an exp
function to control how quickly
weights decrease for more
distant instances. Use in
conjunction with weight By
Distance. Sensible values = 1/5
to 1/10 the number of nearest
neighbors.
WeightByDistance Weight nearest neighbors by their
distance
Class Nominal class, Date class, Missing
class values, numericclass, Binary
class
Attributes Empty nominal attributes, Nominal
attributes, Numeric Attributes,
Unary attributes, Date attributed,
Binary attributes, Missing Values
V. EXPERIMENTAL DATA
A student’s dataset was generated based on the demographic
characteristics, student’s admission dataand pre-collegiate
features of the students. In addition, performance related
measures were also gathered based on class and university
examinations. The study data mining classification algorithms
that are compared in the study includes Naive Bayes, Bayes
Net Classifiers [30], and OneR, J48 decision tree algorithm
which is an open source Java implementation of C4.5
algorithm [31], IBK, JRip algorithm [32] and J48 algorithms.
The information used in this study was collected from college
students enrolled in Bachelor degree program at a 3 reputed
arts and Science College in the state of Tamil Nadu affiliated
to Thiruvalluvar University in the year 2014.The total number
of student’s data was 610 students with 21 attributes were
collected through questionnaire. The collected data was
organized in Microsoft Excel sheet.

ISSN: 2278-2419
135
The target variable was Student End Semester Marks (ESM)
which was usually in numeric form in terms of percentage. It
was discretized using pre-processing filters into 4 categories.
The categories of target variable included First Class (Score >
60%), Second Class (45 - 60%), Third Class (36 - 45%), Fail
(< 36%). Each student record had the following attributes
based on student personal data included gender, category of
admission, living location, family size, and family type, annual
income of the family, father’s qualification and mother’s
qualification. The attributes referring to the students’ pre-
college characteristics includedStudents Grade in High School
and Students Grade in Senior Secondary School. The attributes
describing other college features include thebranch of study of
the students, place of stay, previous semester mark, class test
performance, seminar performance, assignment, general
proficiency, class attendance and performance in the laboratory
work. The study was limited to student’s data collected from
three Arts and Science Colleges in Tamil Nadu. The detailed
description of the dataset is provided in Table 6.
Table 6: Description of the attributes used for
Classification
Variables Description Possible Values
Gender Students Sex {Male, Female}
Branch Students Branch {BCA, B.SC, B.COM,
B.A}
Cat Students category {BC, MBC, MSC, OC,
SBC, SC}
HSG Students grade in
High School
{O – 90% -100%, A –
80% - 89%, B – 70% -
79%,
C – 60% - 69%, D – 50%
- 59%, E – 35% - 49%,
FAIL - <35%}
SSG Students grade in
Senior Secondary
{O – 90% -100%, A –
80% - 89%, B – 70% -
79%,
C – 60% - 69%, D – 50%
- 59%, E – 35% - 49%,
FAIL - <35% }
Medium Medium of
instruction
Tamil, English, others
LLoc Living Location
of Student
{Village, Taluk, Rural,
Town, District}
HOS Student stay in
hostel or not
{Yes, No}
FSize student’s family
size
{1, 2, 3, >3}
FType Students family
type
{Joint, Individual}
FINC Family annual
income
{poor, medium, high}
FQual Fathers
qualification
{no-education,
elementary, secondary,
UG, PG, Ph.D}
MQual Mother’s
Qualification
{no-education,
elementary,
secondary, UG, PG, Ph.D.
NA}
PSM Previous
Semester Mark
{First > 60%, Second >45
&<60%, Third >36
&<45%
Fail < 36%}
CTG Class Test Grade {Poor, Average, Good}
SEM_P Seminar
Performance
{Poor , Average, Good}
ASS Assignment {Yes, No}
GP General
Proficiency
{Yes, No}
ATT Attendance {Poor , Average, Good}
LW Lab Work {Yes, No}
ESM End Semester
Marks
{First > 60% , Second
>45 &<60% , Third >36
&<45%,
Fail < 36%}
VI. EXPERIMENTAL SETUP
The main objective of this research is to study the impact of
feature selection techniques on the classification task so that
classification performance can be improved in the prediction of
student performance for the student performance dataset
generated in the study. The classification model was built using
different algorithms like Naive Bayes, BayesNet, OneR, IBK,
JRip and J48. The WEKA application was used for this
purpose. Each classifier is applied for two testing options -
cross validation (using 10 folds and applying the algorithm 10
times - each time 9 of the folds are used for training and 1 fold
is used for testing) and percentage split (2/3 of the dataset used
for training and 1/3 – for testing).The Feature selection
algorithm tries to select those attributes which have greater
impact on their academic status. Feature Selection Algorithms
used in this study are as follows: CfsSubsetEval, Chi-
SquaredAttributeEval, InfoGainAttributeEval, and
ReliefAttributeEval. Table 7 shows the best attributes that have
selected by Feature Selection Algorithms using WEKA
software tool.
Table 7: Reduction of Attributes using Feature selection
algorithm
Feature subset
Algorithm
Attributes No. of
Attributes
Without Feature
Selection
Algorithms
Sex, Branch, Cat,
SSG_Grade,
HSG_Grade, Medium,
LOC, HOS, FSIZE,
FTYPE, FINC,
FQUAL, MQUAL,
PSM, CTG, SEM_P,
ASS, GP, ATT, LW,
ESM
21
CSER
Branch, SSG_Grade,
FINC, PSM, GP,
ATT, ESM
7
CSAER
PSM, Branch, FINC,
ATT, SSG_Grade,
LW, FSTAT, CTG,
Medium, GP, Sex,
ESM
12
IGAER
PSM, Branch, FINC,
LW, ATT, FSTAT,
SSG_Grade, Medium,
10

ISSN: 2278-2419
136
CTG, ESM
RAER
Branch, PSM, FINC,
LW, Medium, CTG,
FSTAT, ESM
8
VII. RESULTS AND DISCUSSIONS
The Performance of this model is highly depends on selection
of best Attributes from the list of attribute used in student data
set used in student data set. The present investigation focuses
on different Feature Selection Algorithm used in data
preprocessing. Effectiveness of the algorithm is presented in
terms of different measures. For assessing the goodness here
Receiver Operating Characteristics (ROC) value can be
used.ROC value is the representation of the tradeoff between
the false positive and false negative rates. F-Measure, which is
another measure for evaluating the effectiveness, is the
harmonic mean of the precision and recall. The evaluation
measures with variations of ROC values and F-Measure are
generated from an Open Source Data mining tool WEKA.
Average F-measure was computed for feature selection
techniques for each of the classification algorithms. F-measure
determines the predictive accuracy of the classifier. Ten-fold
cross validation was used for this purpose. The results are
summarized in Table 8. The average F-measure was also
calculated for without performing feature selection. The results
clearly show that CSER (0.972) has outperformed other
techniques using 7 attributes. RAER has also produced better
F-Measure (0.961) with 8 attributes.
Table 8: Average F-Measure for each feature subset
Feature
Subset
F – Measure Avera
ge
F-
Measu
re
Naiv
e
Bay
es
Bay
es
Net
One
R
IBK JRip J48
Withou
t
feature
Selecti
on
0.94
3
0.92
4
0.91
5
0.95
3
0.93
9
0.94
4
0.936
CSER 0.99
6
0.96
9
0.98
0
0.99
7
0.98
4
0.90
9
0.972
CSAE
R
0.95
6
0.92
5
0.98
3
0.95
6
0.97
1
0.85 0.940
IGAE
R
0.95
9
0.93
6
0.95
1
0.95
2
0.94
6
0.90
7
0.941
RAER 0.95
8
0.92
3
0.98
1
0.97 0.96
2
0.97
6
0.961
A. Without Feature Selection Algorithm (21 Attributes)
In the present study, Classifiers was implemented Without
Feature Selection algorithm (WFS) applied on the data set and
results of the classifiers shown in Table 9. The results reveal
that the True Positive rate is high for the IBK and OneR, while
it is low for the classifier Bayes Net. The Precision is high for
OneR classifier and it is very low the classifier IBK.
Table 9: Classification results for WFS
Classifier
TP
Rate
Precision
F-
Measure
ROC
Area
J48 0.95 0.95 0.944 0.969
Naïve
Bayes
0.98 0.03 0.943 0.998
Bayes
Net
0.94 0.953 0.924 0.987
IBk 0.985 0.029 0.953 0.98
OneR 0.985 0.985 0.915 0.971
JRip 0.97 0.971 0.939 0.745
Figure 1: Results of WFS
B. CfsSubsetEval with Best First Search Algorithm (7
attributes)
The present study implements CfsSubset Evaluator has been
implemented and results are presented in Table 10. The Table
presents that IBK classifier correctly classifies about 99.680%
for the 10-fold cross-validation testing. It also shows that the
True positive rate is high for Bayes Net. The Precision value is
high for Naïve Bayes and low for J48.
Table 10: Classification results for the CfsSubset Evaluator
Classifier
TP
Rate
Precision Recall
F-
Measure
ROC
Area
J48 0.965 0.96 0.965 0.909 0.968
Naïve
Bayes
0.986 0.998 0.996 0.996 0.986
Bayes
Net
0.997 0.996 0.967 0.969 0.987
IBk 0.99 0.99 0.99 0.997 0.984
OneR 0.980 0.980 0.980 0.980 0.983
JRip 0.985 0.985 0.985 0.984 0.987
From the Fig.3, we observe that the classifier Bayes Net and
JRip has the highest ROC value of 0.987 when it had 7
attributes. The generated macro-averaged F-measure could
attain a maximum of 0.997 for the classifier IBk.
C. Chi-SquareAttributeEval and Ranker (12 Attributes)
The present study implements Chi-Square Attribute Evaluator
has presented in Table 11. The results from Table 6 reveal that
the True Positive rate and precision is high for OneR classifier
and low for J48. It can be verified that OneR classifier

ISSN: 2278-2419
137
correctly classifies about 98.314% for the 10-fold cross-
validation testing.
Figure 2: Results of CfsSubset Evaluator
Table 11: Classification results for Chi-Square Attribute
Evaluation
Classifier
TP
Rate
Precision Recall
F-
Measure
ROC
Area
J48 0.865 0.875 0.865 0.85 0.878
Naïve
Bayes
0.955 0.958 0.955 0.956 0.986
Bayes Net 0.921 0.931 0.921 0.925 0.98
IBk 0.955 0.959 0.955 0.956 0.953
OneR 0.983 0.984 0.983 0.983
0.967
JRip 0.972 0.973 0.972 0.971 0.954
Figure 3: Results of Chi-Square Attribute Evaluation
The graph shows that the classifier Naïve Bayes could attain a
highest ROC value and Bayes Net had second highest value
when it had 12 features. So, we deduce that Naïve Bayes has
the optimal dimensionality in the student data set. Among the
classification Algorithm OneR has the maximum F-measure
value.
D. Information Gain Attribute Evaluator and Ranker (10
Attributes)
The present study implements Information gain Attribute
evaluator on the data set in the WEKA environment and the
results are shown in Table 12. Feature selection algorithm
Information Gain Attribute Evaluation found that Naïve Bayes
classifier correctly classifies about 95.930% for the 10 fold
cross-validation testing. The results from Table 12 also reveal
that the True Positive Rate and precision value is high for
Naïve Bayes and while it is low for J48.
Table 12: Classification results for InfoGain Attribute
Evaluation
Classifier
TP
Rate
Precision Recall
F-
Measure
ROC
Area
J48 0.907 0.913 0.907 0.907 0.935
Naïve
Bayes
0.959 0.96 0.956 0.959 0.988
Bayes
Net
0.93 0.949 0.93 0.936 0.972
IBk 0.953 0.953 0.953 0.952
0.972
OneR 0.952 0.952 0.952 0.951 0.945
JRip 0.947 0.947 0.947 0.926 0.945
Figure 4: Results of InfoGain Attribute Evaluator
We observer from Fig.5, that classifier Naïve Bayes has
highest ROC value and F-measure value found to be maximum
in Naïve Bayes with 10 attributes. Therefore, Naïve Bayes can
achieve relatively good performance on classification tasks.
E. Relief Attribute Evaluation and Ranker (8 Attributes)
The result from the Table 13 shows that Classifier OneR
correctly classifies about 98.181% for the 10-fold Cross-
validation testing on the data set and also True positive rate is
high. It also presents that Precision and True Positive Rate is
low for Bayes Net.
Table 13: Classification results for Relief Attribute Evaluation
Classifier
TP
Rate
Precision Recall
F-
Measure
ROC
Area
J48 0.977 0.977 0.977 0.976 0.955
Naïve
Bayes
0.958 0.962 0.958 0.958 0.987
Bayes
Net
0.915 0.938 0.915 0.923 0.969
IBk 0.97 0.972 0.97 0.97 0.98
OneR 0.982 0.982 0.982 0.981 0.966
JRip 0.964 0.965 0.964 0.962 0.963
Graphical Representation of Table 13 is shown in fig. 6. We
observe that Naïve Bayes could attain the maximum ROC

ISSN: 2278-2419
138
value of 0.987 when it had 8 features. And F-measures are
highest for the classifier OneR.
Figure 5: Results of Relief Attribute Evaluation
The results of the comparative study of six different classifiers
carried out against feature subsets generated by the four
different Feature selection Algorithms are shown below in the
form of Table and Graph. In Table 14, we observe that IBK
comes up with quite good rules for characterizing the structure
in data set. IBK has a highest Accuracy (99.680) over a
CfsSubsetEval(CSER). Also Naïve Bayes shows a second
highest Accuracy (98.569) among the six Classifiers.
Furthermore CfsSubsetEval has the highest accuracy for all the
six classifiers than the Chi-Square, InfoGain and Relief
Attribute Evaluator.
Table 14: Accuracy of Classifiers over Feature Selection
Algorithms
Classifier WFS CSER CSAER IGAER RAER
J48 94.974 96.515 86.516 90.697 90.674
Naïve
Bayes
97.989 98.569 95.505 95.930 95.757
Bayes Net 93.969 97.670 92.134 93.023 91.515
IBk 98.492 99.680 95.405 95.255 96.969
OneR 98.442 98.112 98.314 95.254 98.181
JRip 96.984 98.478 97.191 94.674 96.363
Figure 6: Performance of classifiers over a Feature Selection
Algorithms
VIII. CONCLUSION
The research work aims at analysing the impact of feature
selection techniques on the classification task using feature
selection algorithms WFS, CSER, CSAER, IGAER and RAER
and implementing on the student’s dataset collected from three
arts and science colleges. The analysis done on the resultant
reduced data sets yields faster than models built with no feature
selection. The main concentration of this research work is to
classify the student’s performance in the end semester
examination based on their results and personal characteristics.
In this paper, it is applied various rank based feature selection
filters to data sets to identify the best feature selection
algorithm. Based on the results obtained, the performance of
feature selection algorithms CFS Subset Evaluator was found
to be better than the performance of other three feature
selection algorithms. Also, among the classification
algorithms, the work identify that IBK algorithm yields
99.68% which better than Naive Bayes, BayesNet, OneR and
J48 algorithms. The future work will concentrate on the
implementation of a proposed hybrid method by considering
large dataset collected from many institutions.
References
[1] Parneet Kaur,Manpreet Singh,Gurpreet Singh Josan,
“Classification and prediction based data mining
algorithms to predict slow learners in education sector”,
3rd
Int. Conf. on Recent Trends in Computing, Vol 57,
2015, pp. 500-508.
[2] HythemHashim, Ahmed A, Talab, Ali Satty, Samani A,
Talab, “Data Mining Methodologies To Study Students
Academic Performance Using the C4.5 Algorithm”, Int.
Journal on Computational Sciences & Applications,
Vol.5, No.2, 2015, pp. 59-68.
[3] Vaibhav P. Vasani, Rajendra D. Gawali, “Classification
and performance evaluation using data mining
algorithm”, Int. Journal of Innovative Research in
Science, Engineering and Technology, Vol. 3, Issue 3,
2014, pp. 10453-10458.
[4] AnjanaPradeep and Jeena Thomas, ”Predicting College
Students Dropout using EDM Techniques”, Int. Journal
of Computer Applications, Vol.123,No.5,2015, pp. 26-
34.
[5] PoojaThakar, Anil Mehta, Manisha,”Performance
Analysis and Prediction in Educational Data Mining: A
Research Travelogue”, Int. Journal of Computer
Applications, Vol. 110, No.15, Jan 2015, pp.60-68.
[6] Punlumjeak W, Rachburee N, “A Comparative study of
feature selection techniques for classify student
performance”, 7th
Int. Conf. on Information Technology
and Electrical Engineering, 2015, pp.425-429.
[7] Anal Acharya, Devadatta Sinha, “Application of Feature
Selection Methods in Educational Data Mining”, Int.
Journal of Computer Applications, Vol.103,
No.2,2014,pp.34-38.
[8] Komal S,Sahedani, B supriya Reddy, “A Review: Mining
Educational Data to Forecast Failure of Engineering
Students”, Int. Journal of Advanced Research in
Computer Science and Software Engineering, Vol.3,
Issue 12, 2013, pp. 628-635.
[9] Tarik A. Rashid,” Improving on Classification Models of
Multiple Classes through Effectual Processes”, Int.

ISSN: 2278-2419
139
Journal of Advanced Computer Science and
Applications, Vol.6, No.7, 2015, pp.55-62.
[10] Sadaf Fatima salim Attar, Y.C. Kulkarni, “ Precognition
of Students Academic Failure Using Data Mining
Techniques”, Int. Journal of Advanced Research in
Computer and Communication Engineering, Vol.4, Issue
5, 2015, pp. 109-115.
[11] Carlos Marquez-Vera, Cristobal Romero Morales, and
Sebastian Ventura Soto, “Predicting school failure and
dropout by using data mining techniques”, IEEE Journal
of Latin-American Learning Technologies, Vol. 8, No.
1,2013, pp. 7-14.
[12] Sahil P. Karkhanis, Shweta S. Dumbre, “A Study of
Application of Data Mining and Analytics in Education
Domain”, Int. Journal of Computer Applications,
Vol.120, No.22, 2015, pp. 23-28.
[13] Ogunde A.O and Ajibade D.A,”A Data Mining System
for Predicting University Students Graduation Grades
Using ID3 Decision Tree Algorithm”, Journal of
Computer Science and Information Technology, Vol. 2,
Issue 1, 2014, pp. 21-46.
[14] Dinesh Kumar.A,Radhika.V, “A Survey on Predicting
Student Performance”, Int. Journal of Computer Science
and Information Technologies, Vol.5,
No.5,2014,pp.6147-6149.
[15] ArpitTrivedi, “Evaluation of Student Classification Based
On Decision Tree”, Int.Journal of Advanced Research in
Computer Science and Software Engineering, Vol. 4,
Issue 2, 2014, pp. 111-112.
[16] AgrawalBhawana D, GuravBharti B, “Review on Data
Mining Techniques Used For Educational System”, Int.
Journal of Emerging Technology and Advanced
Engineering, Vol. 4, Issue 11, 2014, pp. 325-329.
[17] Cao, Longbing, Data mining and multi-agent integration,
Springer Science & Business Media, 2009.
[18] Tan, Pang-Ning, Michael Steinbach, and VipinKumar,
“Introduction to data mining”, Boston: Pearson Addison
Wesley, Vol. 1, 2006.
[19] Koller, Daphne, and MehranSahami, "Toward optimal
feature selection", 1996.
[20] Guyon, Isabelle, and André Elisseeff, "An introduction to
feature extraction", Feature extraction, Springer Berlin
Heidelberg, 2006, pp.1-25.
[21] Ramaswami. M and Bhaskaran. R, “A Study on Feature
Selection Techniques in Educational Data Mining”,
Journal of Computing, Vol.1, Issue 1, 2009,pp.7-11.
[22] Forman, George, "An extensive empirical study of
feature selection metrics for text classification",The
Journalof machine learning research, Vol.3, 2003, pp.
1289-1305.
[23] Dy, Jennifer G., et al., "Unsupervised feature selection
applied to content-based retrieval of lung image", Pattern
Analysis and Machine Intelligence, IEEE Transactions on
Vol.25, No.3, 2003, pp. 373-378.
[24] Egozi, Ofer, EvgeniyGabrilovich, and Shaul Markovitch,
"Concept-Based Feature Generation and Selection for
Information Retrieval", AAAI, Vol. 8, 2008, pp. 1132-
1137.
[25] Yu, Lei, and Huan Liu, "Redundancy based feature
selection for microarray data", Proceedings of the tenth
ACM SIGKDD international conference on Knowledge
discovery and data mining, ACM, 2004, pp. 737-742.
[26] Bolon-Canedo, Veronica, Noelia Sanchez-Marono, and
Amparo Alonso-Betanzos, "Feature selection and
classification in multiple class datasets: An application to
KDD Cup 99 dataset", Expert Systems with Applications,
Vol. 38, No.5, 2011, pp. 5947-5957.
[27] Lee, Wenke, Salvatore J. Stolfo, and Kui W. Mok,
"Adaptive intrusion detection: A data mining approach",
Artificial Intelligence Review, Vol. 14, No.6, 2000, pp.
533-567.
[28] Saari, Pasi, TuomasEerola, and Olivier Lartillot,
"Generalizability and simplicity as criteria in feature
selection: Application to mood classification in music",
Audio, Speech, and Language Processing, IEEE
Transactions, Vol. 19, No. 6, 2011, pp.1802-1812.
[29] Mark A. Hall, “Correlation-based Feature Selection for
Machine Learning”, PhD Thesis, The University of
Waikato, 1999.
[30] Mehdi Naseriparsa, Amir-MasoudBidgoli and
Tourajvaraee, “A Hybrid Feature Selection Method to
Improve Performance of a Group of Classification
Algorithms”, Int. Journal of Computer Applications,
Vol.69, No. 17, pp. 28-35.
[31] Russel, Stuart, and Peter Norvig, "Artificial Intelligence:
A Modern Approach ",EUA: Prentice Hall, 2003.
[32] Quinlan, J. Ross,C4. 5: Programs for machine learning,
Elsevier, 2014.
[33] Cohen, William W, "Fast effective rule induction",
Proceedings of the twelth international conference on
machine learning, 1995, pp.115-123.

Performance Evaluation of Feature Selection Algorithms in Educational Data Mining

More Related Content

Similar to Performance Evaluation of Feature Selection Algorithms in Educational Data Mining (20)

More from IIRindia (20)

Recently uploaded (20)

Performance Evaluation of Feature Selection Algorithms in Educational Data Mining