An unsupervised feature selection algorithm with feature ranking for maximizing performance of the classifiers

International Journal of Automation and Computing 12(5), October 2015, 511-517
DOI: 10.1007/s11633-014-0859-5
An Unsupervised Feature Selection Algorithm with
Feature Ranking for Maximizing Performance of
the Classifiers
Danasingh Asir Antony Gnana Singh1
Subramanian Appavu Alias Balamurugan2
Epiphany Jebamalar Leavline3
1
Department of Computer Science and Engineering, Bharathidasan Institute of Technology, Anna University, Tiruchirappalli 620024, India
2
Department of Information Technology, K.L.N College of Information Technology, Sivagangai 630611, India
3
Department of Electronics and Communication Engineering, Bharathidasan Institute of Technology, Anna University,
Tiruchirappalli 620024, India
Abstract: Prediction plays a vital role in decision making. Correct prediction leads to right decision making to save the life, energy,
efforts, money and time. The right decision prevents physical and material losses and it is practiced in all the fields including medical,
finance, environmental studies, engineering and emerging technologies. Prediction is carried out by a model called classifier. The
predictive accuracy of the classifier highly depends on the training datasets utilized for training the classifier. The irrelevant and
redundant features of the training dataset reduce the accuracy of the classifier. Hence, the irrelevant and redundant features must be
removed from the training dataset through the process known as feature selection. This paper proposes a feature selection algorithm
namely unsupervised learning with ranking based feature selection (FSULR). It removes redundant features by clustering and eliminates
irrelevant features by statistical measures to select the most significant features from the training dataset. The performance of this
proposed algorithm is compared with the other seven feature selection algorithms by well known classifiers namely naive Bayes (NB),
instance based (IB1) and tree based J48. Experimental results show that the proposed algorithm yields better prediction accuracy for
classifiers.
Keywords: Feature selection algorithm, classification, clustering, prediction, feature ranking, predictive model.
1 Introduction
Prediction plays a vital role in decision making. This
is done by the predictive model that is known as classi-
fier or supervised learner[1]
. This predictive model is em-
ployed to predict the unseen and unknown data in vari-
ous applications[2]
. Improving the predictive accuracy is
a major challenge among the researchers. The accuracy
of the classifier highly relies on the training dataset which
is used to train the predictive model. The irrelevant and
redundant features of the training dataset reduce the pre-
dictive accuracy[3]
. Hence, irrelevant and redundant fea-
tures must be eliminated from the training dataset for im-
proving the accuracy of the predictive model. The per-
formance improvement of the predictive model include im-
proving the predictive accuracy, reducing the time taken to
build the predictive model and reducing the number of fea-
tures present in the training dataset[3]
. Selecting significant
features from the training dataset for improving the perfor-
mance of the supervised or unsupervised learner is known
as feature selection. This feature selection is classified as
wrapper, filter, embedded and hybrid methods[4]
.
The filter method selects the important features and re-
moves the redundant and irrelevant features from the given
Regular Paper
Manuscript received November 4, 2013; accepted April 3, 2014
Recommended by Editor-in-Chief Huo-Sheng Hu
c Institute of Automation, Chinese Academy of Science and
Springer-Verlag Berlin Heidelberg 2015
dataset using mathematical statistical measure. It performs
well regardless of the classifier chosen, therefore it can work
for any type of classification algorithm[5, 6]
.
The embedded method uses the training phase of the su-
pervised learning algorithm for selecting the important fea-
tures from the training dataset. Therefore, its performance
is based on the classifiers used[7]
.
The wrapper method uses the classifier to determine im-
portant features available in the dataset[8]
. The combi-
nation of filter and wrapper method is known as hybrid
method[4, 9]
.
Many feature selection algorithms are proposed recently,
which concentrate only on removing the irrelevant features.
They do not deal with removing the redundant features
from the training dataset. These redundant features reduce
the accuracy of the classifiers. This paper proposes a filter
based feature selection algorithm named as unsupervised
learning with ranking based feature selection (FSULR).
This algorithm selects the most significant features from the
dataset and removes the redundant and irrelevant features.
The unsupervised learner learns the training dataset using
expectation maximization function and groups the features
into various clusters and these clustered features are ranked
based on the χ2
measure within the cluster. The significant
features from each cluster are chosen based on the threshold
function.

512 International Journal of Automation and Computing 12(5), October 2015
The remaining paper is organized as follows. The rele-
vant literature is reviewed in Section 2. Section 3 discusses
the proposed feature selection algorithm. Section 4 explores
and discusses the experimental results. Conclusion is drawn
in Section 5 with future enhancement of this work.
2 Related work
This section discusses various types of feature selection,
cluster and classification algorithms as a part of the related
works of this proposed algorithm.
2.1 Feature selection algorithm
The feature selection algorithm selects the most signif-
icant features from the dataset using ranking based tech-
niques, subset based and unsupervised based techniques.
In the ranking based technique, the individual features “fi”
are ranked by applying one of the mathematical measures
such as information gain, gain ratio, on the training dataset
“TD”. The ranked features are selected as significant fea-
tures for learning algorithm by a threshold value “TV ” cal-
culated by the threshold function. In subset based tech-
nique, the features of the training datasets are separated
into maximum number of possible feature subsets “S” and
each subset is evaluated by an evaluation criteria to identify
the significance of the features present in the subset. The
subset containing most significant features is considered as
a selected candidate feature subset. In unsupervised based
technique, the cluster analysis is carried out to identify the
significant features from the training dataset[10−12]
.
2.1.1 Feature selection based on correlation (FS-
Cor)
In this feature subset selection, the entire feature set
F = {f1, f2, · · · , fx} of a training dataset “TD” is sub di-
vided into feature subsets “FSi”. Then, two types of the
correlation measures are calculated on each feature subset
“FSi”. One is the feature-feature correlation that is the
correlation measure among the features present in a fea-
ture subset“FSi”, and the other one is feature-class cor-
relation that is the correlation measure between the indi-
vidual feature and the class value of the training dataset.
These two correlation measures are computed for all the in-
dividual feature subsets of the training dataset. The signif-
icant feature subset is identified based on the comparison
between the feature-feature correlation and feature−class
correlation. If the feature−class correlation value is higher
than the feature−feature correlation value, the correspond-
ing feature subset is selected as a significant feature subset
of the training dataset[13]
.
2.1.2 Feature selection based on consistency mea-
sure (FSCon)
In this feature subset selection, entire feature set “F”
of the training dataset “TD” is subdivided into maximum
possible combinations of feature subsets “FSi” and the con-
sistency measure is calculated for each subset to identify the
significant feature subset. The consistency measure ensures
that the same tuple “Ti” is not present in a different class
“Ci”[5, 14, 15]
.
2.1.3 Feature selection based on χ2
(FSChi)
This is a ranking based feature selection technique. The
χ2
statistical measure is applied on the training dataset
“TD” to identify the significance level of each feature
presents in the training dataset. The χ2
value “CV ” is
computed based on the sum of ratio of the difference be-
tween the observed (oi,j) and expected (eij ) frequencies of
the features fi, fj to the expected frequencies (eij) of the
features fi, fj of the possible instance value combinations
of the features[16]
.
2.1.4 Feature selection based on information gain
(FSInfo)
In this feature selection technique, the information gain
measure is applied on the training dataset to identify the
significant features based on information gain value of the
individual features[10]
in terms of entropy. The entropy
value of each feature of the training dataset “TD” is calcu-
lated and ranked based on the information gain value[17]
.
2.1.5 Feature selection based on gain ratio (FSGai-
ra)
In this feature selection technique, the information gain
ratio GR(f) is calculated for each feature of the training
dataset “TD” to identify the significant feature based on
the information present in the features of the “TD”[17]
.
2.1.6 ReliefF
This feature selection technique selects the significant
features from the training dataset “TD” based on the
weighted probabilistic function w(f) with nearest neighbor
principle. If the nearest neighbors of a tuple “T” belong to
the same class, it is termed as “nearest hit”. If the nearest
neighbors of a tuple “T” belong to a different class, it is
termed as “nearest miss”. The probabilistic weight func-
tion value w(f) is calculated based on “nearest hit” and
“nearest miss” values[18−20]
.
2.1.7 Feature selection based on symmetric uncer-
tainty (FSUnc)
This technique uses the correlation measure to select the
significant feature from the training dataset “TD”. In addi-
tion to that, the symmetric uncertainty “SU” is calculated
using the entropy measure to identify the similarity between
the two features fi and fj
[21, 22]
.
2.1.8 Feature selection based on unsupervised
learning algorithm
Unsupervised learning is formally known as clustering
algorithm. This algorithm groups similar objects with re-
spect to the given criteria like density, distance, etc. The
objects present in a group are highly similar than the out-
liers. This technique is applied for selecting the significant
features from the training dataset. Each unsupervised algo-
rithm has its own advantages and disadvantages that deter-
mine the application of each algorithm. This paper utilizes
expectation maximization (EM) clustering technique to se-
lect the significant features by identifying the independent
features in order to remove the redundant features from a
training dataset “TD”[23−25]
.

D. A. A. G. Singh et al. / An Unsupervised Feature Selection Algorithm with Feature Ranking for· · · 513
2.2 Supervised learning algorithm
The supervised learning algorithm builds the predictive
model “PM” by learning the training dataset “TD” to pre-
dict the unlabeled tuple “T”. This model can be built by
various supervised learning techniques such as tree based,
probabilistic and rule based. This paper uses the supervised
learners namely naive bayes (NB), instance based IB1 and
C4.5/J48 supervised learners to evaluate and compare the
performance of the proposed feature selection algorithm in
terms of predictive accuracy and time taken to build the
prediction model with the existing algorithms[26−28]
.
2.2.1 Naive Bayes (NB) classifier
This classifier works based on the Bayesian theory. The
probabilistic function is applied on training dataset “TD”
that contains the features F = {f1, f2, · · ·, fx}, tuple T =
{t1, t2, · · ·, ty} and classes C = {c1, c2, · · ·, cz}. The un-
labeled tuple “UT” is predicted to identify its class Ci
with the probabilistic condition P(Ck|T) > P(Cw|T) where
k = w[29]
.
2.2.2 Decision tree based C4.5/J48 classifier
This tree based classifier constructs the predictive model
using decision tree. Basically, the statistical tool informa-
tion gain is used to learn the dataset and construct the
decision tree. Information gain is computed for each at-
tribute present in the training dataset “TD”. The feature
with higher information value is considered as the root node
and the dataset is divided into further levels. The infor-
mation gain value is computed for all the nodes and this
process is iterated until a single class value is obtained in
all the nodes[30, 31]
.
2.2.3 IB1
This classifier utilizes the nearest neighbor principle to
measure the similarity among the objects to predict the
unlabeled tuple “UT”. The Euclidean distance measure is
used to compute the distance between various tuples present
in the training dataset[32]
.
3 Unsupervised learning with ranking
based feature selection (FSULR)
This algorithm follows a sequence of steps to select the
significant features from the training dataset. Initially, the
training dataset is transposed. Then features are clustered
and the features in each cluster are ranked based on the
χ2
value. Then the threshold value is computed to select
highly significant features from each clustered features. All
the selected features from different clusters are combined
together as candidate significant features selected from the
training dataset.
Algorithm 1. FSULR
Input: Training dataset TD, tuple set T =
{t1, t2, · · · , tl}, feather set F = {f1, f2, · · · , fm}, class la-
bels C = {C1, C2, · · · , Cn} and T, F, C ∈ TD
Output: Selected significant feature subset SSF=
{f1,2 , · · · , fi} ∈ F
Method:
1) Begin
2) Swap (TD)
3) {Return → Swapped TD = SD }// Swap the tuple – T
(row) into feature – F (column)
4) EM Cluster (SD)
5) {Return # of grouped feature sets GF =
{GF1, GF2, GF3, · · ·, GFn}}//GF ∈ F
6) for (L = 1, L ≤ n, L + +)
7) {χ2
(GFL, C)
8) {Return ranked feature members of grouped feature
set RGFK , where K = 1, · · · , n with χ2
value CV }}//
Computing and ranking the χ2
value for the grouped fea-
tures
9) Threshold(CV Max, CV Min)//CV Max, CV Min
are maximum and minimum χ2
values, respectively
10) {Return the value “ϕ”} //ϕ is break-off threshold
value
11) for (M = 1, M ≤ n, M + +)
12) { Break off Feature (RGFM,CVM , ϕ)
13) { Return BFSSK, where K = 1, · · · , n}} //BFSSK
is the break off feature subset based on CV
14) for (N = 1, N ≤ n, N + +)
15) {Return SSF=Union(BFSSN )}//SSF is the se-
lected significant feature subset
16) End.
Phase 1. In this phase, the FSULR algorithm re-
ceives the training dataset “TD” as input with the feature
F = {f1, f2, · · · , fx}, tuples T = {t1t2, · · ·, ty} and class
C = {c1, c2, · · ·, cz}. Then the class C is removed. The
function Swap(·) swaps the features “F” into tuples “T”
and returns swapped dataset “SD” for grouping the fea-
tures “fi”.
Phase 2. In this phase, expectation maximization max-
imal likelihood function[33]
and Gaussian mixture model
as seen in (1) and (2) are used to group the features us-
ing the function EM-Cluster(·). It receives the swapped
dataset “SD” to group the “SD” into grouped feature sets
GF = {GF1, GF2, GF3, · · ·, GFn}.
η(b|Dz) = P(Dz|b) =
m
k=1
P(zk|b) =
m
k=1 l∈L
P(bk|L, b)P (L|b) (1)
where D denotes the dataset containing an unknown pa-
rameter b and known product Z and L is the unseen data.
Expectation maximization is used as distribution function
and this is derived as Gaussian mixture model multivariate
distribution P(z|L, b) as expressed in (2).
P(z|L, b) =
1
(2π)
π
2 Δ l
e− 1
2
(z−λl)Γ −1
l
(z−λl)
. (2)
With this scenario, a value α is randomly picked from
the set “S” with instances “I” from non-continuous set
N = {1, 2, 3, · · · , n}. Let λl be the mean vector of the un-
seen variable and the covariance matrix l(l = 1, 2, · · · , n)

with the discrete distribution P(l|b). The cross-validation
computes the likelihood in correspondence to the value n
to determine n Gaussian components for computing the ex-
pectation maximization. This developed Gaussian mixture
model[34]
is used for clustering the features based on the
level of Gaussian component possessed by the individual
features “fi”[35]
and return the n number of grouped fea-
ture subsets GF = {GF1, GF2, GF3, · · ·, GFn}.
Phase 3. In this phase, the most significant features are
selected from the clustered feature subset by χ2
(·) function.
It receives the clustered feature set GFK with the class fea-
ture C and computes the χ2
value “CV ” for all the features
as shown in (3)[16]
CV =
m
i=1
n
j=1
(oij − eij )
eij
. (3)
The feature “fi” contains distinct tuple values
tv1, tv2,· · ·, tvn. Then, the χ2
value “CV ” is computed for
the features fi and fj . oij is the observed frequency and
eij is the expected frequency of possible tuple values. To
identify the significant features, the features are ranked by
their χ2
value “CV ” and the ranked grouped feature set is
obtained. The Threshold(·) function receives the CV Min
and CV Max values and returns the break-off threshold
value ϕ as shown in (4). The function Break off Feature()
recognizes the ranked grouped feature set RGF with ϕ and
breaks off the top ranked “CV ” up to the value ϕ and
returns them as break off feature subset BFSS.
ϕ = H ×
α − β
2
+ β (4)
where ϕ is break off-threshold value, H is threshold constant
that is equal to 1, α is the maximum χ2
value (CV Max)
of the feature, and β is the minimum χ2
value (CV Min)
of the feature.
Phase 4. This phase combines break off feature subset
BFSS threshold from the clustered features by the func-
tion Union(·) and produces the selected significant feature
subset SSF as the candidate selected features.
4 Experimental setup
Totally, 21 datasets are used to conduct the experiment
with number of features ranging from 2 to 684, number of
instances from 5 to 1728 and number of class labels from 2
to 24 as listed in the Table 1. These datasets are taken from
the Weka software[36]
and UCI Repository[37]
. The perfor-
mance of the proposed algorithm FSULR is analyzed and
compared with the other feature selection algorithms FS-
Cor, FSChi, FSCon, FSGai-ra, ReliefF, FSUnc and FSInfo
with NB, J48 and IB1 classifiers.
5 Experimental procedure, results and
discussion
This experiment is conducted with 21 well known publi-
cally available datasets as shown in the Table 1 by 7 feature
selection algorithms namely FSCor, FSChi, FSCon, FSGai-
ra, ReliefF, FSUnc and FSInfo. Initially, the datasets are
fed as an input to the feature selection algorithms and the
selected features are obtained as output of the feature se-
lection algorithms. These selected features are given to the
Table 1 Datasets
Serial number Dataset Number of features Number of instances Number of classes
1 Breast cancer 9 286 2
2 Contact lenses 24 5 5
3 Credit–g 20 1000 2
4 Diabetes 8 768 2
5 Glass 9 214 7
6 Ionosphere 34 351 2
7 Iris 2D 2 150 3
8 Iris 4 150 3
9 Labor 16 57 2
10 Segment challenge 19 1500 7
11 Soybean 683 36 14
12 Vote 435 17 3
13 Weather nominal 4 14 2
14 Weather numeric 14 5 2
15 Audiology 69 226 24
16 Car 6 1728 4
17 Cylinder bands 39 540 2
18 Dermatology 34 366 3
19 Ecoli 7 336 8
20 Anneal 39 898 6
21 Hay-train 10 373 9

NB, J48 and IB1 classifiers and the predictive accuracy and
the time taken to build the predictive model are calculated
with the 10-fold cross validation test mode.
The performance of the proposed FSULR algorithm is
analyzed in terms of predictive accuracy, time to build the
model and the number of features reduced. Fig. 1 expresses
the overall performance of feature selection algorithms in
producing the predictive accuracy for supervised learners
and it is observed that the FSLUR performs better than
all the other feature selection algorithms compared. The
second and third positions are retained by the FSCon and
FSInfo respectively. Fig. 2 shows that the proposed FSULR
achieves better accuracy than the other algorithms com-
pared for NB and IB1 classifiers. The FSCon achieves bet-
ter results for J48 classifier as compared to all the other
classifiers.
Fig. 1 Comparison of overall prediction accuracy in percentage
with respective feature selection algorithm
Fig. 2 Comparison of prediction accuracy in percentage with
respect to the feature selection algorithm and classifier
From Fig.3, it is evident that FSULR takes more time
to build the predictive model compared to all other algo-
rithms. From Fig.4, it is observed that the FSGai-ra and
FSUnc require less time to build the model for NB classi-
fier than other algorithms compared. The FSUnc takes less
time to build model for J48 classifier. The FSGai-ra and
ReliefF consume less time to build the predictive model for
IB1 classifier. Fig.5 exhibits that ReliefF reduces the num-
ber of features significantly than other algorithms compared
and FSCon reduces the least number of features.
Fig. 3 Comparison of overall time taken to build the predictive
model (in second) with respect to the feature selection algorithm
Fig. 4 Comparison of time taken to build the predictive model
(in second) with respect to the feature selection algorithm and
classifier
Fig. 5 Comparison of number of features reduction with respect
to the feature selection algorithm
6 Conclusions and future work
This paper proposed a feature selection algorithm namely
unsupervised learning with ranking based feature selection
algorithm (FSULR). Performance of this algorithm is an-

alyzed in terms of accuracy produced for classifier, time
taken to build predictive model and feature reduction. The
performance of this algorithm is compared with other fea-
ture selection algorithms namely correlation based FSCor
χ2
based FSChi, consistency based FSCon, gain ratio based
FSGai-ra, ReliefF, symmetric uncertainty based FSUnc and
information gain based FSInfo algorithm with NB, J48, IB1
classifiers. The FSULR achieves better prediction accuracy
than the other feature selection algorithms and achieves
higher accuracy for IB1 and NB classifiers compared to the
other feature selection algorithms. The FSULR is also con-
siderably good in reducing the time to build model for IB1
compared to FSGai-ra and ReliefF. FSULR is considerably
good in reducing the time to build model for J48 classi-
fiers compared to FSCon algorithm. The FSULR reduces
the number of features compared to FSCor, FSCon and
FSGai-ra. In future, this work can be extended with other
statistical measures for ranking the features within the clus-
ters.
References
[1] J. Sinno, Q. Y. Pan. A survey on transfer learning. IEEE
Transactions on Knowledge and Data Engineering, vol. 22,
no. 10, pp. 1345–135, 2010.
[2] M. R. Rashedur, R. M. Fazle. Using and comparing different
decision tree classification techniques for mining ICDDR, B
Hospital Surveillance data. Expert Systems with Applica-
tions, vol. 38, no. 9, pp. 11421–11436, 2011.
[3] M. Wasikowski, X. W. Chen. Combating the small sam-
ple class imbalance problem using feature selection. IEEE
Transactions on Knowledge and Data Engineering, vol. 22,
no. 10, pp. 1388–1400, 2010.
[4] Q. B. Song, J. J. Ni, G. T. Wang. A fast clustering-based
feature subset selection algorithm for high-dimensional
data. IEEE Transactions on Knowledge and Data Engineer-
ing, vol. l5, no. 1, pp. 1–14, 2013.
[5] J. F. Artur, A. T. Mário. Efficient feature selection filters for
high-dimensional data. Pattern Recognition Letters, vol. 33,
no. 13, pp. 1794–1804, 2012.
[6] J. Wu, L. Chen, Y. P. Feng, Z. B. Zheng, M. C. Zhou,
Z. H. Wu. Predicting quality of service for selection by
neighborhood-based collaborative filtering. IEEE Transac-
tions on Systems, Man, and Cybernetics: Systems, vol. 43,
no. 2, pp. 428–439, 2012.
[7] C. P. Hou, F. P. Nie, X. Li, D. Yi, Y. Wu. Joint embedding
learning and sparse regression: A framework for unsuper-
vised feature selection. IEEE Transactions on Cybernetics,
vol. 44, no. 6, pp. 793–804, 2014.
[8] P. Bermejo, L. dela Ossa, J. A. Gámez, J. M. Puerta.
Fast wrapper feature subset selection in high-dimensional
datasets by means of filter re-ranking Original Research
Article. Knowledge-based Systems, vol. 25, no. 1, pp. 35–44,
2012.
[9] S. Atulji, G. Shameek, V. K. Jayaramanb. Hybrid biogeog-
raphy based simultaneous feature selection and MHC class
I peptide binding prediction using support vector machines
and random forests. Journal of Immunological Methods,
vol. 387, no. 1-2, pp. 284–292, 2013.
[10] H. L. Wei, S. Billings. Feature subset selection and ranking
for data dimensionality reduction. IEEE Transactions on
Pattern Analysis and Machine Intelligence, vol. 29, no. 1,
pp. 162–166, 2007.
[11] X. V. Nguyen, B. James. Comments on supervised fea-
ture selection by clustering using conditional mutual
information-based distances. Pattern Recognition, vol. 46,
no. 4, pp. 1220–1225, 2013.
[12] A. G. Iffat, S. S. Leslie. Feature subset selection in large di-
mensionality domains. Pattern Recognition, vol. 43, no. 1,
pp. 5–13, 2010.
[13] M. Hall. Correlation-based Feature Selection for Machine
Learning, Ph. D dissertation, The University of Waikato,
New Zealond, 1999.
[14] Y. Lei, L. Huan. Efficient feature selection via analysis of
relevance and redundancy. Journal of Machine Learning Re-
search, vol. 5, no. 1, pp. 1205–1224, 2004.
[15] M. Dash, H. Liu, H. Motoda. Consistency based feature se-
lection. In Proceedings of the 4th Pacific Asia Conference
on Knowledge Discovery and Data Mining, Kyoto, Japan,
pp. 98–109, 2000.
[16] H. Peng, L. Fulmi, C. Ding. Feature selection based
on mutual information criteria of max-dependency, max-
relevance, and min-redundancy. IEEE Transactions on
pp. 1226–1238, 2005.
[17] H. Uˇguz. A two-stage feature selection method for text cat-
egorization by using information gain, principal component
analysis and genetic algorithm. Knowledge-based Systems,
vol. 24, no. 7, pp. 1024–1032, 2011.
[18] M. Robnik-ˇSikonja, I. Kononenko. Theoretical and empir-
ical analysis of ReliefF and RReliefF. Machine Learning,
vol. 53, no. 1-2, pp. 23–69, 2003.
[19] S. Yijun, T. Sinisa, G. Steve. Local-learning-based feature
selection for high-dimensional data analysis. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence, vol. 32,
no. 9, pp. 1610–1626, 2010.
[20] W. Peng, S. Cesar, S. Edward. Prediction based on in-
tegration of decisional DNA and a feature selection algo-
rithm RELIEF-F. Cybernetics and Systems, vol. 44, no. 3,
pp. 173–183, 2013.

[21] H. W. Liu, J. G. Sun, L. Liu, H. J. Zhang. Feature selec-
tion with dynamic mutual information. Pattern Recogni-
tion, vol. 42, no. 7, pp. 1330–1339, 2009.
[22] M. C. Lee. Using support vector machine with a hybrid fea-
ture selection method to the stock trend prediction. Expert
Systems with Applications, vol. 36, no. 8, pp. 10896–10904,
2009.
[23] P. Mitra, C. A. Murthy, S. K. Pal. Unsupervised feature
selection using feature similarity. IEEE Transactions on
pp. 301–312, 2002.
[24] J. Handl, J. Knowles. Feature subset selection in unsu-
pervised learning via multi objective optimization. Inter-
national Journal of Computational Intelligence Research,
vol. 2, no. 3, pp. 217–238, 2006.
[25] H. Liu, L. Yu. Toward integrating feature selection algo-
rithms for classification and clustering. IEEE Transactions
on Knowledge and Data Engineering, vol. 17, no. 4, pp. 491–
502, 2005.
[26] S. Garc´ıa, J. Luengo, J. A. Sáez, V. Lóez, F. Herrera. A
survey of discretization techniques: Taxonomy and empir-
ical analysis in supervised learning. IEEE Transactions on
Knowledge and Data Engineering, vol. 25, no. 4, pp. 734–
750, 2013.
[27] S. A. A. Balamurugan, R. Rajaram. Effective and efficient
feature selection for large-scale data using Bayes theorem.
International Journal of Automation and Computing, vol. 6,
no. 1, pp. 62–71, 2009.
[28] J. A. Mangai, V. S. Kumar, S. A. alias Balamurugan.
A novel feature selection framework for automatic web
page classification. International Journal of Automation
and Computing, vol. 9, no. 4, pp. 442–448, 2012.
[29] H. J. Huang, C. N. Hsu. Bayesian classification for data
from the same unknown class. IEEE Transactions on Sys-
tems, Man, and Cybernetics, Part B: Cybernetic, vol. 32,
no. 2, pp. 137–145, 2002.
[30] S. Ruggieri. Efficient C4.5. IEEE Transactions on Knowl-
edge and Data Engineering, vol. 14, no. 2, pp. 438–444,
2002.
[31] P. Kemal, G. Salih. A novel hybrid intelligent method based
on C4.5 decision tree classifier and one-against-all approach
for multi-class classification problems. Expert Systems with
Applications, vol. 36, no. 2, pp. 1587–1592, 2009.
[32] W. W. Cheng, E. Hüllermeier. Combining instance-based
learning and logistic regression for multilabel classification.
Machine Learning, vol. 76, no. 3, pp. 211–225, 2009.
[33] J. H. Hsiu, P. Saumyadipta, I. L. Tsung. Maximum likeli-
hood inference for mixtures of skew Student-t-normal dis-
tributions through practical EM-type algorithms. Statistics
and Computing, vol. 22, no. 1, pp. 287–299, 2012.
[34] M. H. C. Law, M. A. T. Jain, A. K. Jain. Simultaneous fea-
ture selection and clustering using mixture models. IEEE
Transactions on Pattern Analysis and Machine Intelligence,
vol. 26, no. 9, pp. 1154–1166, 2004.
[35] T. W. Lee, M. S. Lewicki, T. J. Sejnowski. ICA mix-
ture models for unsupervised classification of non-gaussian
classes and automatic context switching in blind signal sep-
aration. IEEE Transactions on Pattern Analysis and Ma-
chine Intelligence, vol. 22, no. 10, pp. 1078–1089, 2002.
[36] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reute-
mann. The WEKA data mining software: An update. ACM
SIGKDD Explorations Newsletter, vol. 11, no. 1, pp. 10–18,
2009.
[37] K. Bache, M. Lichman. UCI Machine Learning Repository,
[Online], Available: http://archive.ics.uci.edu/ml, Irvine,
CA: University of California, School of Information and
Computer Science, 2013.
Danasingh Asir Antony Gnana Singh
received the M. Eng. and B. Eng. degrees
from Anna University, India. He is cur-
rently working as a teaching fellow in the
Department of Computer Science and En-
gineering, Bharathidasan Institute of Tech-
nology, Anna University, India.
His research interests include data min-
ing, wireless networks, parallel computing,
mobile computing, computer networks, image processing, soft-
ware engineering, soft computing, teaching learning process and
engineering education.
E-mail: asirantony@gmail.com (Corresponding author)
ORCID iD: 0000-0002-4913-2886
Subramanian Appavu Alias Balamu-
rugan received the Ph. D. degree from
Anna University, India. He is currently
working as a professor and the head of the
Department of Information Technology in
K.L.N College of Information Technology,
India.
His research interests include pattern
recognition, data mining and informatics.
E-mail: app s@yahoo.com
Epiphany Jebamalar Leavline re-
ceived the M. Eng. and B. Eng. degrees
from Anna University, India, and received
the MBA degree from Alagappa Univer-
sity, India. She is currently working as
an assistant professor in the Department of
Electronics and Communication Engineer-
ing, Bharathidasan Institute of Technology,
Anna University, India.
Her research interests include image processing, signal pro-
cessing, VLSI design, data mining, teaching learning process and
engineering education.
E-mail: jebi.lee@gmail.com

An unsupervised feature selection algorithm with feature ranking for maximizing performance of the classifiers

More Related Content

What's hot (17)

Viewers also liked (17)

Similar to An unsupervised feature selection algorithm with feature ranking for maximizing performance of the classifiers (20)

Recently uploaded (20)

An unsupervised feature selection algorithm with feature ranking for maximizing performance of the classifiers