Optimization Technique for Feature Selection and Classification Using Support Vector Machine

INTERNATIONAL JOURNAL FOR TRENDS IN ENGINEERING & TECHNOLOGY
VOLUME 4 ISSUE 1 – APRIL 2015 - ISSN: 2349 - 9303
13
Optimization Technique for Feature Selection and
Classification Using Support Vector Machine
Abstract— Classification problems often have a large number of features in the data sets, but only some of them are
useful for classification. Data Mining Performance gets reduced by Irrelevant and redundant features. Feature selection
aims to choose a small number of relevant features to achieve similar or even better classification performance than using
all features. It has two main objectives are maximizing the classification performance and minimizing the number of
features. Moreover, the existing feature selection algorithms treat the task as a single objective problem. Selecting attribute
is done by the combination of attribute evaluator and search method using WEKA Machine Learning Tool. We compare
SVM classification algorithm to automatically classify the data using selected features with different standard dataset.
Index Terms— Data Mining, Kernel methods, Support Vector Machine, WEKA, Classification.
.
——————————  ——————————
1 INTRODUCTION
upport Vector Machine (SVM) was first described in 1992,
introduced by Boser, Guyon, and Vapnik. Support vector
machines (SVMs) are a set of related supervised learning
methods used for classification and regression [1]. They belong to a
family of linear classifiers. In other words, Support Vector Machine
(SVM) is a tool of classification and regression prediction that make
use of the machine learning theory to increase the predictive
accuracy by automatically avoiding the data over-fit. Support Vector
machines can also be defined as systems that uses the hypothesis
space of linear functions in a multi dimensional feature space,
trained with a learning protocols from optimization theory that
implements a learning bias derived from the statistical learning .
SVM becomes famous when, using pixel maps as input; it gives
accuracy comparable to sophisticated neural networks with
elaborated features in a handwriting recognition task [2]. It is also
being used for many applications, such as hand writing, face, texture
analysis and so on, especially for classification and regression based
applications. The fundamentals of Support Vector Machines (SVM)
has been developed by Vapnik [3] and gained popularity due to many
promising features such as better empirical performance. SVM
attempts to minimize a bound on the generalization error (i.e., the
error made by the learning machine on test data not used during
training).
2 FEATURE SELECTION
Feature selection (also known as subset selection or variable
selection) is a process commonly employed in machine learning to
solve the high dimensionality problem. It select a subset of important
features and removes irrelevant and redundant features for simpler
and more concise data. The advantages of feature selection are multi-
fold. First, feature selection greatly saves the running time of a
learning process by removing irrelevant, redundant and noisy
features. Second, without the interaction of irrelevant and redundant
features, learning algorithms can focus on the most important
perspective of the data and build accurate data models. Thus, the
classification performance can be improved. Third, feature selection
can help us to build a simpler and more generic model and provide a
better insight into the underlying concept of the task.
2.1 Feature Selection Objectives
Different feature selection algorithms may have various
objectives to achieve. The following is a list of common objectives
used by researchers:
1. Find the minimally sized feature subset that is necessary and
sufficient to the target concept.
2. Select a subset in a set of N features from a set of M features,
N < M, such that the value of a criterion function is optimized over
all subsets of size N.
3. Choose a subset of features for improving prediction accuracy or
decreasing the size of the structure without significantly decreasing
prediction accuracy of the classifier built using only the selected
features.
4. Select a small subset such that the resulting class distribution,
which provides the values of the selected features, is as close as
possible to the class distribution for all given feature values.
2.2 Feature Selection Procedure
The process begins with subset generation that employs a
certain search strategy to produce feature subsets. Then each subset
is evaluated according to an evaluation criterion and compared with
the previous best result. If the evaluated subset is better, then it
replaces the previous one, else remains the same. The process of
subset generation and evaluation gets repeated until it satisfies the
given criteria. Figure 1 show the Feature Selection procedure.
S
M. Kumar1
S. Chidambaram M.E2
Dr. K G. Srinivasagan3
1M.E. (student) 2
Asst.Professor, Dept of IT 3
Professor & Head, Dept of CSE (PG)
National Engineering College National Engineering College National Engineering College
kumarec92@gmail.com chidambaramraj1@gmail.com kgsnec@rediffmail.com

14
Finally the selected best subset is validated by using the prior
knowledge or some test data. Search strategy and evaluation criteria
are two key points in the study of feature selection.
Fig 1. Feature Selection Procedure
1 SUBSET GENERATION
Subset generation begins with a search start point, which can be
of any one of the following an empty set, the full set, or a randomly
generated subset. From the starting point, it can search feature
subsets in different directions, such as forward, backward, and
random. In forward search, features gets added at a time, while in
backward search the least important feature is removed based on
evaluation criterion. Random search adds or deletes features at
random to avoid being trapped into local maxima. There are various
search strategies for finding an optimal or suboptimal feature subset.
2 EVALUATION CRITERIA
After the generation of feature subsets, the results are evaluated
by certain criteria to measure their goodness. Generally, the goodness
of feature subsets means that the discriminating ability of subsets
that distinguish among or within different classes. Based on their
dependent in the learning algorithms, this algorithms can be broadly
classified into Wrapper Method, Filter Method. Wrapper method is
used to find feature subsets with high accuracy because the features
match well with the learning algorithms.
3. STOPPING CRITERIA
A feature selection process gets terminated under any one of the
following criteria:
i. Whether the search is complete.
ii. Whether a predefined size of feature subsets is selected.
iii. Whether a predefined number of iterations are executed.
iv. Whether an optimal or sufficiently good feature subset
according to the evaluation function has been obtained.
v. Whether the change (addition or deletion of features) of feature
subsets does not produce a better subset.
4. RESULT VALIDATION
In some applications, the relevant features are known
beforehand. Then we can able to validate the feature selection results
by comparing the prior knowledge. However, in the real-world
applications we do not know which features are relevant.
We have to use the classification performance on test data as an
indicator of the goodness of the selected feature subsets.
Once the feature selection is done the selected features are used
for classification. In Attribute Selection module irrelevant features
are discard and the selected feature is used for classification in order
to improve better performance. Feature Selection procedure is shown
in the figure 2. WEKA Machine learning tool is used for feature
selection and classification for the given datasets.
Attribute selection involves searching through all possible
combination of attributes in the data to find which subset of attribute
work best for improving performance of classification. WEKA has a
very flexible combination of search and evaluation methods for the
dataset’s attributes. It consists of Search methods and Attribute
Evaluator.
Fig 2. Attribute Selection
2.3 ATTRIBUTE EVALUATOR
The Attribute Evaluator is the method by which a subset of
attributes is assessed. For example, they may be assessed by building
a model and evaluating the accuracy of the model.
ClassifierSubsetEval is used as an attribute evaluator to access the
attributes of given datasets. It Evaluates attribute subsets on training
data or a separate hold testing set.
Option in ClassifierSubsetEval
•classifier -- Classifier to use for estimating the accuracy of subsets
•holdOutFile -- File containing hold out/test instances.
•useTraining -- Use training data instead of hold out/test instances.
2.4 SEARCH METHOD
The Search Method is the is the structured way in which the
search space of possible attribute subsets is navigated based on the
subset evaluation. Search method traverses the attribute space to find
a good subset. Quality of the subset is measured by the attribute
subset evaluator. BestFirst search method is used to find the possible
search space of the attributes.
BestFirst searches the space of attribute subsets by greedy hill
climbing augmented with a backtracking ability. Setting the number
of consecutive non-improving nodes will allow the level of
backtracking. Best first may start with the empty set of attributes and
proceed forward, or start with the full set of attributes and proceed
backward, or start at any point and proceed in both the direction. The
combination of best first search method and ClassifierSubsetEval
will give the best feature selection in result which is further used for
improving classification accuracy.
Option in BestFirst
Attributes
Subset
Generation
Subset
Evaluation
Result
Validation
Stopping
Criterion
NO
YES
ATTRIBUTE EVALUATOR SEARCH METHOD

15
•Direction -- Set the direction of the search.
•LookupCacheSize -- Set the maximum size of the lookup cache of
evaluated subsets. This can be expressed as the multiplier of the
given number of attributes in the given data set. (default = 1).
•Search Termination -- Set the amount of backtracking.
3 LIBSVM CLASSIFICATION
A wrapper class is one of the libsvm tools (the libsvm classes
that is the jar file need to be in the path of the class to use this
classifier).LibSVM executes faster than SMO since it uses LibSVM
to develop the SVM classifier. LibSVM allows users to do
experiment with One-class, Regressing , and nu-SVM supported by
LibSVM tool. LibSVM reports many useful information about
LibSVM classifier (e.g., confusion matrix, precision, recall, ROC,
etc.).
The goals of SVM are to separate the data along with its hyper
plane and extend ed this to non-linear boundaries using kernel trick
[8] [5]. For calculating the SVM we see that the goal is to correctly
classify all the data. For mathematical calculations we have,
[a] If Zi= +1; (1)
[b] If Zi= -1; (2)
[c] For all Zi; (3)
In this equation x is a vector point and w is weight and is also a
vector. So to separate the data [a] should always be greater than that
of zero. Among all possible hyper planes, SVM selects the one plane
where the distance of that hyper plane should be larger as possible.
If the training data is good then the every test vector will locate in
radius r from training vector. Hence the chosen hyper plane is
located at the farthest distance from the data. This desired hyper
plane that maximizes the margin also bisects the lines between
closest points on convex hull of the two datasets.
Distance from closest point on hyper plane to origin can be
found by maximizing the value of x where x is on that hyper plane.
Similarly for the other side points we have a same scenario. Thus by
solving and subtracting the two distances we get the summed
distance from the separated hyper plane to the nearest points.
Maximum Margin is
M = 2 / ||v|| (4)
3.1 Kernel Trick
If data is linear, then to divide the data the separating hyper plane
is used. However it shows that the data is far away from linear and
the datasets are mostly inseparable. Thus the kernels are used to map
the non-linear input data to a high-dimensional space. It shows that
new mapping is then linearly separable.
K(xi,xj) = φ(xi) T
φ(xj ) (5)
Feature Space: Transforming the data into feature space makes it
possible to define a similarity measure on the basis of the dot
product. If the space of the feature is chosen suitably,then pattern
recognition can be easy [1].
Fig 3. Representation of Support Vectors
(4a) Linearly inseparable (4b) Linearly separable
The Kernel trick allows SVM’s to form boundaries of nonlinear.
Steps followed in kernel trick are given below.
[a] The algorithm is expressed using the data sets inner products.
This is also knows as dual problem.
[b] Original data are traveled through nonlinear maps to get new
data with respect to new dimensions by adding a pair wise product of
some of the original data dimension to each data vector.
[c] Rather than an inner product on these new, large vectors, and
placed in tables and later do a lookup table, we can provide a dot
product after doing nonlinear mapping on them of the data. This
function is called as kernel function. More kernel functions are as
fallows.
The kernel function plays a critical role in SVM and also on its
performance. It is based upon reproduction Spaces of Kernel Hilbert
[8].Then such kernel represents a legitimate product of inner in the
space of feature. The training set is not separable linearly in a space of
input. The training set is separable linearly in the space of feature. This
is called as the ―Kernel trick‖.
The different kernel functions are listed below where r --coef0
value coef0 for polynomial kernel -r --coef0 value coef0 for sigmoid
kernel. d --degree value degree for polynomial kernel. g --gamma
value gamma for polynomial kernel. g --gamma value gamma for rbf-
kernel. g --gamma value gamma for sigmoid kernel default 0 is set for
all gamma values. u & v are kernel parameter user defined.
1 bvxi
1 bvxi
1 bvxi
+
+ 0 0 0
+ + +
+ 0 0 0
0 0
+ 0 0 0
+ + 0
+ + +

16
TABLE I. KERNEL METHODS
4 EXPERIMENT AND RESULT
This section describes the experimental results obtained by
applying the proposed algorithms to a variety of data sets. For
experimentation, five benchmark datasets are taken from the UCI
machine learning repository [9] as shown in Table 2. In order to
validate the prediction results of the comparison of the classification
techniques and the 10-fold crossover validation is used. The present
study divided the data into 10 folds where 1 fold was for testing and
9 folds were for training for the 10-fold crossover validation.
TABLE II. PROVIDES THE ATTRIBUTE INFORMATION OF FIVE DATASETS
Datasets Features Instances Class
Iris 5 150 3
Ecoli 8 336 8
Mushroom 23 8124 2
Diabetes 9 768 2
Breast cancer 10 286 2
WEKA is a collection of machine learning algorithms used for
data mining tasks. The algorithms can either be directly applied to a
dataset or called into the java code.
WEKA contains tolls for data pre-processing the data,
classification of data, regression, data clustering, association rules
and visualization.
WEKA (Waikato Environment for Knowledge Analysis) is a
data mining system developed by the University of Waikato in New
Zealand that implements data mining algorithms. WEKA is the art
facility for developing machine learning (ML) techniques and their
application to real-world data mining problems. The algorithms are
applied directly to a dataset.UCI data sets are used for both feature
selection and classification using LibSVM in WEKA and their
performance is compared.
ROC curve is plot by selecting the visualize threshold cure in
WEKA classifier result buffer. It is one of the performance metrics
for the classification algorithm to know its efficient prediction.
TP rate, FP rate, Precision, recall, F-measure and ROC area are
the other performance metrics to know the LibSVM performance.
These metrics are used to form the confusion matrix.
4.1Performance Metrics
Three classical evaluation metrics of Precision data, Recall and
F-score are used to validate the efficiency of the proposed method.
The three metrics are traditionally defined as a classification of
binary task with both positive and negative classes. Precision is the
positive predictions proportion that are correct, and recall is the
positive samples proportion that are correctly positively predicted.
That is:
Precision = 𝑇𝑃/ (𝑇𝑃+𝐹𝑃)
Recall = 𝑇𝑃/ (𝑇𝑃+𝐹𝑁)
F = score = 2 ∗ 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 ∗ 𝑅𝑒𝑐𝑎𝑙𝑙Precision + Recall
 True positive (TP) = number of positive samples correctly
predicted.
 False negative (FN) = number of positive samples wrongly
predicted.
 False positive (FP) = number of negative samples wrongly
predicted as positive.
 True negative (TN) = number of negative samples correctly
predicted.
TABLE III. ACCURACY VALUES AND ERROR VALUES FOR ORIGINAL
DATASETS USING LIBSVM CLASSIFIER
Dataset
Correctly
classified
Instances
(Accuracy %)
Incorrectly
classified
Instances
(Error %)
Iris 96.66 3.333
Ecoli 75.5 24.40
Mushroom 99.803 0.1669
Diabetes 65 34
Breast cancer 70 29
Table3 shows the various datasets with correctly and incorrectly
classified instances after applying LibSVM algorithm to the data.
Firstly, SVM has nonlinear dividing hyperplane, which give high
discrimination. Secondly, SVM provides good generalization ability
for unseen data classification these properties of SVM used to solve
the non-linear problem more efficiently.
Kernel name Representation
Linear u'*v
Polynomial (gamma*u'*v+ coef0)^degree
Gaussian(RBF)
exp(-gamma*|u-v|^2)
Sigmoid
tanh(gamma*u'*v + coef0)

17
TABLE IV. RESULT FOR DATASETS USING LIBSVM CLASSIFIER
Table4 shows the LibSVM classification result for different data sets.
5 CONCLUSION
This paper proposed a SVM based optimization algorithm,
which can used to optimize the parameter SVM values and also used
to obtain the features of optimal subset by applying LIBSVM
method to remove insignificant features and effectively find best
parameter values.
The goal of this paper is to design Support Vector Machine
Algorithm to find the classification accuracy. Feature Selection
algorithm is used to improve the classification accuracy of classifier
with respect to datasets.
REFERENCES
[1] Support-vector networks. Corinna Cortes and Vladimir
Vapnik, Machine Learning, 1995, Volume 20, Number 3,
Pages 273-297.
[2] Yuan-HaiShao, Wei-JieChen , Jing-JingZhang “An efficient
weighted Lagrangian twin support vector machine for
imbalanced data classification” journal homepage
:www.elsevier.com/locate/pr.
[3] Winda Astuti, Rini Akmeliawati, Wahju “Hybrid Technique
Using Singular Value Decomposition (SVD) and Support
Vector Machine(SVM) Approach for Earthquake Prediction”
IEEE Journal Of Selected Topics In Applied Earth
Observations And Remote Sensing, Vol. 7, No. 5, May 2014.
[4] Kanu Patel, Jay Vala, Jaymit Pandya “Comparison of various
classification algorithms on iris datasets using WEKA”
International journal of Advance Engineering and Research
Development (IJAERD) Volume 1 Issue 1, February 2014,
ISSN: 2348 – 4470.
[5] J.Platt “Fast training of SVMs using sequential minimal
optimization”, In B. Sch¨olkopf, C.Burges and A.Smola (ed.),
Advances in Kernel Methods: Support Vector Learning, MIT
Press, Cambridge, MA, 1999, 185-208.
[6] O.L.Mangasarian and D.R.Musicant “Lagrangian support
vector machines”, Journal of Machine Learning Research, 1,
2008, 161-177.
[7] Z.Nematzadeh Balagatabi1, H.Nematzadeh Balagatabi
“Comparison of Decision Tree and SVM Methods in
Classification of Researcher’s Cognitive Styles in Academic
Environment” Indian Journal of Automation and Artificial
Intelligence Vol: 1 Issue: 1 January 2013 ISSN 2320 – 4001.
[8] Y.Tang,Y.-Q.Zhang,N.V.Chawla,S.Krasser,‖SVMS modeling for
highly imbalanced classification‖, IEEE
Trans.Syst.ManCybern.B:Cybern.39(1) (2009) 281–288.
[9] G. Fung and L.Mangasarian “Proximal support vector
machine classifiers”, Proceedings of the 7th ACM conference
on knowledge discovery and data mining, ACM, 2001, 77-86.
[10] Y.Lee, O.Mangasarian, W.Wolberg, “Survival-time
classification of breast cancer patients”, Computational
Optimization and Applications, 25(2003), 151-166.
[11] Z.Q. Qi, Y.J. Tian, Y. Shi,” Robust Twin Support Vector
Machine for Pattern Classification”, Pattern Recognition,
2013, 46(1): 305-316.
[12] Y. Tian, Y. Shi, X. Liu, “Recent advances on support vector
machines research”, Technological and Economic
Development of Economy 18(1) (2012), 5-33.
[13] Z. Qi, Y. Tian, and Y. Shi, “Structural Twin Support Vector
Machine for Classification, Knowledge-Based Systems, 2013,
DOI:10.1016/j.knosys.2013.01.008.
[14] S. Maldonado, R. Weber, J. Basak,” Kernel-penalized SVM for
feature selection”, Inform. Sci. 181 (2011) 115–128.
[15] L. Song, A. Smola, A. Gretton, J. Bedo, K. Borgwardt,
“Feature selection via dependence maximization”, J. Mach.
Learn. Res. 13 (2012) 1393–1434.
[16] G. Victo Sudha George, V. Cyril Raj, “Review on feature
selection techniques and the impact of svm for cancer
classification using gene expression profile”, Int. J. Comput.
Sci. Eng. Surv. 2 (3) (2011) 16–27.
[17] H. Yu, J. Kim, Y. Kim, S. Hwang, Y.H. Lee, “An efficient
method for learning nonlinear ranking SVM functions”,
Inform. Sci. 209 (2012) 37–48.
[18] S. Maldonado, J. López, “Imbalanced data classification using
second-order cone programming support vector machines”,
Pattern Recogn. 47 (2014).
[19] E. Carrizosa, B. Martín-Barragán, D. Romero-Morales,
“Detecting relevant variables and interactions in supervised
classification”, Euro. J. Oper. Res. 213 (2011) 260–269.
[20] R. Hassan, R.M. Othman, P. Saad, S. Kasim, “A compact
hybrid feature vector for an accurate secondary structure
prediction”, Inform. Sci. 181 (2011)5267–5277.
Dataset TP rate FP rate Precision Recall F-measure ROC Area
Iris 0.967 0.017 0.967 0.967 0.967 0.975
Ecoli 0.756 0.087 0.622 0.756 0.681 0.613
Mushroom 0.998 0.002 0.996 0.998 0.998 0.998
Diabetes 0.651 0.651 0.424 0.651 0.513 0.500
Breast cancer 0.706 0.688 0.695 0.706 0.594 0.509

Optimization Technique for Feature Selection and Classification Using Support Vector Machine

More Related Content

What's hot (13)

Similar to Optimization Technique for Feature Selection and Classification Using Support Vector Machine (20)

More from IJTET Journal (20)

Recently uploaded (20)

Optimization Technique for Feature Selection and Classification Using Support Vector Machine