On Feature Selection Algorithms and Feature Selection Stability Measures : A Comparative Analysis

International Journal of Computer Science & Information Technology (IJCSIT) Vol 9, No 3, June 2017
DOI:10.5121/ijcsit.2017.93014 159
ON FEATURE SELECTION ALGORITHMS AND
FEATURE SELECTION STABILITY MEASURES: A
COMPARATIVE ANALYSIS
Mohana Chelvan P1
and Dr. Perumal K2
1
Dept. of Computer Science, Hindustan College of Arts and Science,Chennai – 603 103,
India,
2
Dept. of Computer Applications, Madurai Kamaraj University, Madurai – 625 021,
India,
ABSTRACT
Data mining is indispensable for business organizations for extracting useful information from the huge
volume of stored data which can be used in managerial decision making to survive in the competition. Due
to the day-to-day advancements in information and communication technology, these data collected from e-
commerce and e-governance are mostly high dimensional. Data mining prefers small datasets than high
dimensional datasets. Feature selection is an important dimensionality reduction technique. The subsets
selected in subsequent iterations by feature selection should be same or similar even in case of small
perturbations of the dataset and is called as selection stability. It is recently becomes important topic of
research community. The selection stability has been measured by various measures. This paper analyses
the selection of the suitable search method and stability measure for the feature selection algorithms and
also the influence of the characteristics of the dataset as the choice of the best approach is highly problem
dependent.
KEYWORDS
Data Mining, Feature Selection, Feature Selection Algorithms, Selection Stability, Stability Measures.
1. INTRODUCTION
Data mining is essential for getting useful information from huge amount of data stored due to the
day-to-day activities of the organizations. These data are mostly high dimensional which makes
the data mining task difficult. Feature selection is a scheme which chooses small related feature
subsets from the dataset. Feature selection improves accuracy, efficiency and model
interpretability of the algorithms. The subsets get by feature selection on the same sample should
be similar in subsequent iterations and should be stable even for small perturbations or the
addition of new data. Feature selection stability is the robustness of feature selection algorithms
for small perturbations in the dataset. Otherwise it will create confusion in researcher’s mind
about the result and lowers their confidence in their conclusion of research work [1]. Recently
selection stability becomes hot topic of research. This paper gives an account of various selection
stability measures and their application in various feature selection algorithms.
2. FEATURE SELECTION ALGORITHMS
The feature selection process is mostly based on three approaches viz. filter, wrapper and hybrid
[2]. The filter approach of feature selection is by removing features on some measures or criteria
and the feature’s wellness is examined using intrinsic or statistical features of the dataset. A

160
feature is referred as a much suited feature based on these properties, and is chosen for machine
learning or data mining uses. In the wrapper approach the subset of features is produced and then
goodness of the subset is examined with the use of a classifier. The purpose of some classifier
here is ranking the features of the dataset depending on which an option is chosen for the desired
use. The embedded model combines the advantages of both the above models. The hybrid
approach takes benefits of both the approaches by using the various examination criterions of
them in various search stages. The review on feature selection algorithms is presented in [3] and
the important feature selection algorithms are shown below.
2.1. One-R
One-R algorithm is put-forth by Holte [4] and is simple. The algorithm has each rule for every
aspect in the training data and fixes the rule with minimum error. This algorithm considers the
mathematically charged features as continuous. This is one of the most primitive techniques. It
just separates the series of values to many dis-joint intervals also it is a straightforward method. It
treats missing values as a legitimate value called “missing”. Here simple rules are produced
depending on one feature only. It can be helpful to determine a standard performance as a target
for more learning techniques even though it is a minimal form of classifier.
2.2. Information Gain (IG)
The entropy is a criterion of impurity in a training set S. It is defined as a means which provide
extra data about Y presented by X that shows the value by which the entropy of Y falls [5]. This
scheme is termed as IG which is symmetrical in nature and is given in (1).
IG = H(Y) − H(Y/X) = H(X) − H(X/Y) (1)
Knowledge obtained regarding Y followed by the observation of X equals the knowledge obtained
regarding X followed by observation of Y. This is oriented in support of features with more values
though it may be less informative and is the weakness of the IG criterion. It computes the worth
of an attribute by considering the information gain based on the class as in (2). By considering the
variation between the entropy of the feature and the conditional entropy provided the class label,
the IG decides the independence within a feature and the class label.
IG (Class, Attribute) = H (Class) −H (Class | Attribute) (2)
2.3. Gain Ratio (GR)
The GR is a non-symmetrical measure. It is coined to make compensation for the IG bias [5]. GR
is contributed in (3).
IG
GR = (3)
H(X)
IG is normalized by dividing using the entropy of X, and vice versa when the variable Y has to be
determined. The range of GR values lie within the span of [0, 1] because of the normalization. A
value of GR = 0 represents the absence of relation among Y and X while GR = 1 shows that the
information about X easily predicts Y. The GR favors variables with less values as opposed to IG.
2.4. Symmetrical Uncertainty (SU)
The Symmetric Uncertainty is given in (4). By dividing IG by the sum of the entropies of X and
Y, the Symmetrical Uncertainty criterion compensates for the inherent IG bias [5].

161
IG
SU = 2 (4)
H(Y) + H(X)
The values taken by SU are normalized to the range [0, 1] due to the correction factor 2. A value
of SU = 0 indicates that X and Y are uncorrelated while the value of SU = 1 predicts that
information about one feature generally depicts the other. The SU is biased toward features with
fewer values as GR.
2.5. Correlation-based Feature Selection (CFS)
The attribute subsets value is evaluated by CFS by considering the degree of redundancy between
them together with the distinct predictive capability of each feature. Subsets of features which are
greatly correlated with the class with less inter-correlation are given the preference [6]. CFS
determines the optimal feature subset and can be brought together with other search mechanisms.
Authors have GA as search method with CFS as fitness function. CFS is given in (5).
k rzi
rzc = (5)
√ k + (k − 1) rii
where rzc gives the correlation within the added feature subsets and the class variable, k refers to
the number of subset features, rzi is the mean correlations within the subset features and the class
variable, and rii is the mean inter-correlation within subset features [6].
2.6. ReliefF
To evaluate the attributes by assigning a weight to every feature, the feature capacity to
differentiate within the classes has been used. Using sampling repetition to sample an occurrence,
with the value of particular feature considered for the adjacent occurrence of similar and various
classes, the ReliefF attribute contribution [7] examines the value of feature. Relevant features are
then selected based on which whose weights cross a user meaned threshold. The weight
determination depends on the possibility of two adjacent neighbors belonging to the same class
with similar value of the feature and the possibility of the adjacent neighbors from two various
classes with various values for a feature. The feature is much important during occurrence of
much variation within these probabilities. The measure is described for a two-class problem.
However, by parting the challenge to a two class problems series, it could also be expanded to
handle various classes. ReliefF trial to increase the margin where it is connected to
hypothesis margin maximization. Relief offers a good trade-off between complexity and
performance.
2.7. ChiSquare
Like Information Gain, ChiSquare computes the independency among a feature and a class-
label using variation within the feature’s entropy and the Conditional entropy provide the class-
label. ChiSquare analyses if a specific nature does not depend on the class-label [8]. The 2 way
Chi-squared test is an analytical technique which identifies closeness of the expected and the
actual outcomes. This technique makes an assumption that the parameters are chosen randomly
and obtained from a suitable sample of independent values. The resultant Chi-squared values
show the deviation of results from the random (expected) outcome. This technique computes the
absence of independency within a term and the category. Chi-Squared is the general analytical
exam which computes deviation from the distribution anticipated if someone assume the feature

162
happening does not depend on the class value. Being an analytical exam, it shows error values for
few small anticipated counts that are normal in classification of test due to presence of rare
happening word features and sometimes due to possession of small positive training instances for
a theme. Here, the χ2 test is used to examine the independency of both events, where the two
events A and B are said to be not dependent only if the condition P(AB) = P(A)P(B) occurs or,
its equivalent, P(A|B) = P(A) and P(B|A) = P(B). In feature selection, the two events are term and
class occurrence. Feature selection by the χ2 statistic is similar like doing a hypothesis exam on
the class distribution because it connects to the feature query values.
The Null Hypothesis shows absence of co-relation; every value can have instance in a particular
class like other classes. Consider on Null Hypothesis, if p instances has a particular value and q
instances occur in a particular group, (p·q)/n instances have a particular value and are in a
particular class. This is since p/n instances with the value and q/n instances are in the class, and if
probabilities don’t depend, then (i.e. the null hypothesis) their total probability is given by their
multiple. Provided the Null Hypothesis, the χ2 statistic computes deviation of actual value
compared to the expected one.
r c (Oi,j − Ei,j)2
x2
= ∑ ∑ (6)
i=1 j=1 Ei,j
Here in (6), ‘r’ is the number of various values of the feature in examination, ‘c’ gives the number
of classes in examination (in this work, c = 2), O i,j is the number of instances having value i
present in class j, and Ei,j is the anticipated instances number having value I and class j,
depending upon (p·q)/n. The greater this Chi-squared statistic, the more un-likely the
independency the values and classes distribution; i.e., if they are connected, the feature in
examination favours the class.
2.8. Fisher
The Fisher score is sometimes coined as the information score, as this shows the quantity of
knowledge which is provided by a variable about a unknown variable on whom it is based [9].
Computation of the score is done by computing the deviation of the knowledge of observed one
from the expected one. When variance is reduced, information is increased. As the score
expectation is zero, the Fisher information also gives the score variance. Fisher score gives larger
scores to the feature which discriminates different samples from different classes easily.
3. SELECTION STABILITY MEASURES
The stability measures have been divided to three main categories depending on the
representation of the output of the selection method [10]. The categories are stability by index,
stability by rank and stability by weight. Let A and B are features subset, A, B ⊂ X, of the similar
dimension i.e., cardinality, k. Let r = |A ∩ B| be the cardinality of the two subset intersections.
The desirable properties that each stability measurement should have [11] are as follows:
• Monotonicity. For a fixed number of features, n and subset size, k, the higher the
intersection within the subsets, the larger the stability index value.
• Limits. The index must be destined by constants that don’t rely on n or k. When the two
subsets are identical, maximum value should be attained, i.e., for r = k.
• Correction for chance. The index must be a perpetual value to draw independent feature
subsets of similar cardinality, k.

163
In addition to these requirements, the following important properties should be taken into
consideration due to their impact on the selection stability result [1] [12].
• The size of the dataset.
• The amount of chosen features.
• The sample size.
• The data variance.
• The symmetry of the measurement.
4. CATEGORIES OF FEATURE SELECTION STABILITY MEASURES
Stability can be assessed by the pairwise comparison between the resulting subsets obtained by
feature selection algorithm on datasets. The stability is higher if the similarity between the
resulting subsets is greater. Based on the output of the feature selection technique, the stability
measures are of three different representations i.e., indexing, ranking, and weighting [10].
4.1. Stability by Index
In this category of measurements, the selected subset of features is signified as a binary vector
with cardinality equivalent to the total features m or as a vector of indices relating to the selected
features k. Unlike the other stability measurements i.e., rank or weight based measurements, the
index measurements have the possibility for handling subset of feature, i.e., the number of
selected features k ≤ m. The index measurements assess the amount of overlap between the
resulting subsets of features for assessing the stability. The examples for stability by index
measurement are Dice’s Coefficient, Tanimoto Distance, Jaccard Index and Kuncheva Index.
4.2. Stability by Rank
The stability by rank method assesses the stability by evaluating the correlation between the
ranking vectors. Unlike the index method, these methods do not deal with partial set of features as
they cannot tackle vectors with multiple cardinality i.e., vectors that resemble to various features
set. The measurements in this category include Spearman’s Rank Correlation Coefficient SRCC.
4.3. Stability by Weight
Similar to the stability by rank, this category of measurement deals with only full subset of
features. This method assesses selection stability by evaluating the weight of the full feature set.
The stability by weight category of measurement has only one member called the Pearson's
Correlation Coefficient PCC. Here the stability is assessed by evaluating the correlation between
the two sets of weights wi and wj for the whole feature dataset.
5. IMPORTANT FEATURE SELECTION STABILITY MEASURES
5.1 Dice's Coefficient
Dice, Tanimoto and Jaccard are similar index based stability measures. Dice coefficient calculates
selection stability by computing the overlap among two subsets of features as in (7) and is used in
[5].
2│Ƒ'1∩Ƒ'2│
Dice (Ƒ'1, Ƒ'2) = (7)
│Ƒ'1│+│Ƒ'2│

164
Dice bounds between the values of 0 and 1, where 0 means the results are unstable i.e., no overlap
between the subset of features and 1 means the two subsets are stable or identical.
5.2 Jaccard Index (JI)
The given different results R = {R1, R2... Rl} will correspond to l different folds of the sample
dataset. By evaluating the overlapping instances between the subsets in R, the stability can be
evaluated as in (8). By estimating the number of overlap among the features of the chosen
subsets, the JI is to assess the stability for feature subset with selected feature indices [13]. The
similarity between finite numbers of subsets is measured by the Jaccard coefficient measures. JI is
measured as the intersection dimension of the selected feature subsets divided by the size of their
union. JI for two selected subsets is shown by (8) and for a number of subsets in subsequent
iterations is shown by (9).
|Ri ∩ Rj |
SJ (Ri, Rj) = (8)
|Ri U Rj |
2 l-1 l
SJ(R) = ∑ ∑ SJ (Ri, Rj ) (9)
l (l-1) i=1 j=i+1
The Jaccard Index SJ gives a value which bounds within the interval [0, 1] where 0 defines the
two subsets Ri and Rj of feature selection consequences are not steady and overlapped and 1
defines the results are very steady and identical.
5.3 Kuncheva Index (KI)
The stability index is an index based measure based on correction for chance and cardinality of
the intersection. KI obey all the requirement appeared in [11] and is the only measurement as
such.
│Ƒ’1 ∩ Ƒ’2│. n – k2
KI (Ƒ’1, Ƒ’2) = (10)
k ( n – k)
In (10), k is the total features of the subset and n is the total features of the dataset. KI's results
ranges in [–1, 1], where –1 means there is no intersection between the lists and k = n/2. KI
achieves 1 when Ƒ'1 and Ƒ'2 are same to define the intersection set cardinality equals k. KI values
becomes nearer to zero for drown individually lists. KI becomes desirable measure because of the
correction for chance term which was introduced in [11]. Unlike other measurements, larger value
of cardinality will not affect the stability in KI. In case of the other stability measures, the larger
the cardinality is, the higher the stability will be.
5.4 Spearman's Rank Correlation Coefficient (SRCC)
SRCC is rank based stability measure and is shown in (11) to evaluate the stability of two ranked
sets of features' r and r’. It is introduced by A. Kalousis et al. in [10].
(rt − r‫׳‬t)2
SRCC(r, r‫׳‬) = 1 – 6 ∑ (11)
t m (m2
− 1)

165
The result of Spearman's will be in the range of [–1, 1]. The maximum will be achieved when the
two ranks are identical while the minimum is when they are exactly in inverse order and 0 mean
no correlation at all between r and r‫׳‬.
5.5 Pearson's Correlation Coefficient (PCC)
Pearson's is weight based stability measure and is used to measure the correlation between the
weights of the features that returned from more than one run and is adapted in [10]. The Pearson's
Correlation Coefficient PCC stability will be as in (12).
∑i (wi − µw) (w‫׳‬I − µw')
PCC (w, w‫)׳‬ = (12)
√∑i (wi − µw)2
∑i (w‫׳‬I − µw')2
Here µ is the mean. PCC takes values between –1 and 1, where 1 means the weight vectors are
perfectly correlated, –1 depicts they are anti correlated while 0 means no correlation. The stability
will be shown higher when the weight is equal to zero for big number of features. However, this
will not be an issue in situations as the algorithm assigns weight between 1 and −1. The PCC is a
symmetric measure. It is the only stability measure that handles feature weights.
6. EXPERIMENTAL RESULTS
The two datasets used in the experiments are Census-Income (KDD) dataset and Insurance
Company Benchmark (COIL 2000) dataset. The datasets are obtained from the KEEL dataset
repository [14]. The dataset characteristics are shown in Table 1.
Table 1. Characteristics of the datasets census and coil 2000.
Filter model is mostly stable, very efficient and highly generalizable because it is independent of
any classifier [7]. Due to these advantages, most of the analysis in this paper belongs to this
model. However it might not be accurate as wrapper model. Well-known Filter algorithms used in
the experiment include One-R [9], Information Gain [5], Gain Ratio [5], Symmetric Uncertainty
[5], CFS [10], ReliefF [11], ChiSquare [12] and Fisher [13]. All algorithms are filter-based; hence
classifier is not included in any selection process. The search methods include BestFirst and
Ranker.
The BestFirst method examines the attribute space subsets using greedy hillclimbing augmented
with a backtracking capability. The BestFirst method may search backward after beginning with
the full set of attributes or search forward after beginning with the empty set of attributes or it
may search in both directions by making use of all suitable single attribute additions and deletions

166
at a specific point after starting at any point. The search method BestFrst is well suited for CFS,
ChiSquare and Fisher. The Ranker method ranks attributes with particular estimations. It will be
used in combination with attribute evaluators One-R, Information Gain, Gain Ratio, Symmetric
Uncertainty and ReliefF. The table 2 summarizes the result of the experiments.
In the experiments, dataset census is mostly categorical while the dataset coil 2000 is completely
numerical. Some feature selection algorithms are better suited for the dataset census while others
are better suited for the dataset coil 2000. The dataset census is consistent for all feature selection
algorithms except CFS. But the results with the dataset coil 2000 have some fluctuations. The
algorithms ReliefF, Information Gain and Symmetric Uncertainty give good results with coil
2000 dataset while CFS gives the worst result.
Table 2. Selection stability results for feature selection algorithms for the datasets census and coil 2000.
The index based selection stability measures i.e., Dice’s Coefficient, Jaccard Index and Kuncheva
Index can work with partial set of features while the rank based and weight based selection
stability measures i.e., SRCC and PCC can work with full set of features. The feature selection
stability measure value will be improved up to the optimum number of selected features and then
decreases and so the number of selected features will be kept at optimum number for index based
measures and is not possible for rank based and weight based measures. The Dice’s coefficient
has slightly higher values of stability measure due to larger value of cardinality of intersection of
selected features i.e., for Fisher in comparing with ChiSquare and CFS. However, the minimum
value of stability measure for Kuncheva Index i.e., for CFS is due to the correction of chance
term. In the case of Kuncheva Index, the larger value of cardinality will not affect selection
stability in comparison with other stability measures. Fig. 1 gives the comparison chart of
selection stability outcomes for the feature selection algorithms for the experimental datasets.

167
Figure 1. Comparison of selection stability results for the feature selection algorithms for datasets census
and coil 2000.
7. CONCLUSION
Feature selection can play a major role in data mining and it has been identified as a challenging
research problem for the academicians, industrialists and researchers. Researchers of feature
selection must understand the underlying sampling techniques because the dataset distribution is
generally un-known. There is really no "best" method for feature selection because different data
sets have different measures of correlation. Even when the fold is sampled from the similar
dataset, with no overlap with other folds, may tend with different related features. The behaviors
of the feature selection algorithm will vary on different datasets with different characteristics. In
addition, selection of suitable stability measure for the feature selection algorithm is also an
interesting research problem.
REFERENCES
[1] Salem Alelyani, Huan Liu, “The Effect of the Characteristics of the Dataset on the Selection
Stability”, 1082-3409/11, IEEE DOI 10.1109/ International Conference on Tools with Artificial
Intelligence.2011.167, 2011.
[2] K. Mani, P. Kalpana, “A review on filter based feature selection”, International Journal of Innovative
Research in Computer and Communication Engineering (IJIRCCE) ISSN: 2320-9801, Vol. 4, Issue 5,
May 2016
[3] K. Sudha, J. JebamalarTamilselvi, "A Review of Feature Selection Algorithms for Data Mining
Techniques", International Journal on Computer Science and Engineering (IJCSE) ISSN: 0975-3397,
Vol. 7, No.6, pp. 63-67, June 2015.
[4] Holte, R.C., “Very simple classification rules perform well on most commonly used datasets”,
Machine Learning, 11: 63-91, 1993.
[5] Hall, M.A., and Smith, L. A., “Practical feature subset selection for machine learning”, Proceedings
of the 21st Australian Computer Science Conference, 181–191, 1998.
[6] Mark A. Hall, ”Correlation-based Feature Selection for Machine Learning”, Dept. of Computer
science, University of Waikato. http://www.cs.waikato.ac.nz/ mhall / thesis.pdf, 1998.
[7] Marko, R.S., and Igor, K., “Theoretical and empirical analysis of relief and reliefF”, Machine
Learning Journal, doi: 10.1023/A: 1025667309714, 53: 23–69, 2003.
[8] I. Witten and E. Frank. Data Mining: Practical machine learning tools and techniques. Morgan
Kaufmann Pub, 2005.
[9] R. Duda, P. Hart, and D. Stork. Pattern Classification. John Wiley & Sons, New York, 2 edition,
2001.
[10] A. Kalousis, J. Prados, and M. Hilario, “Stability of feature selection algorithms: a study on high-
dimensional spaces”, Knowledge and Information Systems, 12(1):95 – 116, May 2007.

168
[11] L. I. Kuncheva, “A stability index for feature selection”, In Proceedings of the 25th conference on
Proceedings of the 25th IASTED International Multi Conference: artificial intelligence and
applications, Anaheim, CA, USA,. ACTA Press, 390 – 395, 2007.
[12] Salem Alelyani, Zheng Zhao, Huan Liu, “A Dilemma in Assessing Stability of Feature Selection
Algorithms”, 978-0-7695-4538-7/11, IEEE DOI 10.1109/ International Conference on High
Performance Computing and Communications. 2011.99, 2011.
[13] Y. Saeys, T. Abeel, and Y. Van de Peer, Robust feature selection using ensemble feature selection
techniques, 2008.
[14] J. Alcala-Fdez, A. Fernandez, J. Luengo, J. Derrac, S. García,L.Sanchez, and F. Herrera, “KEEL data-
mining software tool: Data set repository, integration of algorithms and experimental analysis
framework”, J. Multiple-Valued Logic Soft Comput., 17(2): 255–287, 2010.
AUTHORS
P. Mohana Chelvan working as an Assistant Professor in Department of Computer Science
at Hindustan College of Arts and Science, Chennai, India since 2015. His educational
qualifications are MCA, NIELIT C Level (IT), MPhil. (CS) and UGC NET. He is currently
Ph.D. rese arch scholar in computer science from Madurai Kamaraj University in the area of
Privacy preserving data mining.
Dr. K. Perumal working as an Associate Professor in Department of Computer Applications
at Madurai Kamaraj University, Madurai, India since 1990. He awarded his Ph.D. degree in
computer science from Madurai Kamaraj University in the area of Digital image processing.
He has contri buted more than 40 papers in the International Journals and Conferences and
also editor of proceedings for National Conference on Contemporary Developments in
Information and communication technologies. He has guiding 9 scholars. His research
interest includes Data Mining, Big Data and image processing especially in medical image processing.

On Feature Selection Algorithms and Feature Selection Stability Measures : A Comparative Analysis

More Related Content

What's hot (18)

Similar to On Feature Selection Algorithms and Feature Selection Stability Measures : A Comparative Analysis (20)

More from AIRCC Publishing Corporation (20)

Recently uploaded (20)

On Feature Selection Algorithms and Feature Selection Stability Measures : A Comparative Analysis