Applied Soft Computing 9 (2009) 1–12

Contents lists available at ScienceDirect

Applied Soft Computing

journal homepage: www.elsevier.com/locate/asoc

Review

Dimensionality reduction based on rough set theory: A review

K. Thangavel a, A. Pethalakshmi b,*
a
Department of Computer Science, Periyar University, Salem 636011, Tamil Nadu, India
b
Department of Computer Science, Mother Teresa Women’s University, Kodaikanal 624102, Tamil Nadu, India

A R T I C L E I N F O A B S T R A C T

Article history: A rough set theory is a new mathematical tool to deal with uncertainty and vagueness of decision system
Received 6 August 2006 and it has been applied successfully in all the fields. It is used to identify the reduct set of the set of all
Received in revised form 22 May 2008 attributes of the decision system. The reduct set is used as preprocessing technique for classification of
Accepted 26 May 2008
the decision system in order to bring out the potential patterns or association rules or knowledge through
Available online 14 July 2008
data mining techniques. Several researchers have contributed variety of algorithms for computing the
reduct sets by considering different cases like inconsistency, missing attribute values and multiple
Keywords:
decision attributes of the decision system. This paper focuses on the review of the techniques for
Rough set
Reduct
dimensionality reduction under rough set theory environment. Further, the rough sets hybridization
Neural network with fuzzy sets, neural network and metaheuristic algorithms have also been reviewed. The performance
Metaheuristic analysis of the algorithms has been discussed in connection with the classification.
Knowledge and classification ß 2008 Elsevier B.V. All rights reserved.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2. Feature selection and classification using rough set theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
3. Hybrid approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.1. Feature selection and classification using rough and fuzzy sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.2. Rough sets and neural network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.3. Metaheuristic algorithms with rough set theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
4. Decision system with missing data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
5. Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
6. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Acknowledgements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

1. Introduction
Rough set theory was initially developed [76,77] for a finite
universe of discourse in which the knowledge base is a partition,
which is obtained by any equivalence relation defined on the
universe of discourse. In rough set theory, the data is organized in a
table called decision table. Rows of the decision table correspond to
objects, and columns correspond to attributes. In the data set, a
class label indicates the class to which each row belongs. The class
* Corresponding author. Tel.: +91 451 2460285.
E-mail addresses: drktvelu@yahoo.com (K. Thangavel),
pethalakshmi@yahoo.com (A. Pethalakshmi).
label is called as decision attribute, the rest of the attributes are the
condition attributes. Here, C is used to denote the condition
attributes, D for decision attributes, where C \ D = K, and tj
denotes the jth tuple of the data table. Rough set theory defines
three regions based on the equivalent classes induced by the
attribute values: lower approximation, upper approximation and
boundary. Lower approximation contains all the objects, which are
classified surely based on the data collected and upper approx-
imation contains all the objects, which can be classified probably,
while the boundary is the difference between the upper
approximation and the lower approximation.
Let U be any finite universe of discourse. Let R be any
equivalence relation defined on U. Clearly, the equivalence relation
partitions U. Here, (U, R) is the collection of all equivalence classes,

1568-4946/$ – see front matter ß 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.asoc.2008.05.006
2 K. Thangavel, A. Pethalakshmi / Applied Soft Computing 9 (2009) 1–12
is called the approximation space. Let W1, W2, W3, . . ., Wn be the
elements of the approximation space (U, R). This collection is
known as knowledge base. Then for any subset A of U, the lower
and upper approximations are defined as follows:
 
Wi
RA ¼ [ (1)
Wi  A
 
Wi
R̄A ¼ [
W i \ A 6¼ f
An ordered pair (RA; R̄A) is called a rough set. After defining
these approximations of A, the reference universe U is divided into
three different regions: the positive region POSR(A), the negative
region NEGR(A) and the boundary region BNDR(A), defined as
follows:
POSR ðAÞ ¼ RA
NEGR ðAÞ ¼ U  R̄A
BNDR ðAÞ ¼ R̄A  RA
Hence, it is trivial that if BND(A) = F, then A is exact. This approach
provides a mathematical tool that can be used to find out all
possible reducts. Some of the rough set-based techniques are
available in [84,105].
The main advantage of rough set theory is that it does not need
any preliminary or additional information about data: like
probability in statistics or basic probability assignment in
Dempster–Shafer theory and grade of membership or the value
of possibility in fuzzy set theory.
Several software systems based on rough set theory have been
implemented in the areas of medicine, pharmacology, engineering,
banking, market analysis, conflict analysis, pattern recognition,
environment, linguistics and gene expression. Many real-life and
non-trivial applications of this methodology have also been reported
in the literature [5,10,31,45]. In this paper reduct, dimensionality
reduction and feature selection are used interchangeably.
The rest of the paper is organized as follows. Section 2 discusses
the similarity, dissimilarity, merits and demerits of several
methods for feature selection and classification using rough set
theory. Section 3 describes the hybrid approach for feature
selection and classification. Rough set with missing data is
described in Section 4. Section 5 presents the various real-life
applications. The relationship and performance analysis of the
various algorithms are studied and tabulated in each section.
Section 6 concludes the paper.
2. Feature selection and classification using rough set theory
Recently, researchers have focused their attention on reduct
and classification algorithms based on rough sets [61,87,119].
Feature selection process refers to choosing subset of attributes
from the set of original attributes. The purpose of the feature
selection is to identify the significant features, eliminate the
irrelevant of dispensable features and build a good learning model.
The benefits of feature selection are twofold: it considerably
decreases the computation time of the induction algorithm and
increases the accuracy of the resulting mode. Feature selection has
been studied intensively in the past one decade [4,49,52,68].
A decision table may have more than one reducts. Anyone of
them can be used to replace the original table. Finding all the
reducts from a decision table is NP-Hard [67]. Fortunately, in
many real applications it is usually not necessary to find all of
them. It is sufficient to compute only one reduct [39]. A natural
question is which reduct is the best if there is more than one
reduct. The selection depends on the optimality criterion
associated with the attributes. If it is possible to assign a cost
function to attributes, then the selection can be naturally based
on the combined minimum cost criteria. In the absence of an
attribute cost function, the only source of information to select
the reduct is the contents of the data table [73]. For simplicity,
we adopt the criteria that the best reduct is the one with the
minimal number of attributes and that if there are two or more
(2) reducts with the same number of attributes, then the reduct
with the least number of combinations of values of its attributes
is selected.
Nowadays numerous successful implementations of feature
selection and classification of various applications using rough set
theory are available [46,97,98]. The applications are summarized
in Table 1 and they are discussed in this section.
In many real applications, it was not easy to incorporate low-
(3) level libraries. Due to this, Griffin and Chen introduced an approach
which added a rough set extension to an existing scripting
(4)
language Tcl [23]. This was named as RSL Tcl provided an easy-to-
use interface and had a great deal of flexibility. They developed two
(5)
new Tcl commands which were used to induce rules from decision
tables. The Tcl commands were easily accepted through web pages
and this approach was useful for incremental and automated
construction of knowledge bases.
Khoo et al. proposed a novel approach for the classification and
rule induction of inconsistent information systems [51]. It was
achieved by integrating rough set theory with a statistics-based
inductive learning algorithm. The framework of a prototype rough
set-based classification system (R-class) was also presented. This
R-class technique was compared with the other rule techniques
like ID3 and LERS. For each possible rule generated, R-class was
able to provide an estimation of the expected classification
reliability. This assisted users in deciding the appropriateness of
the rules generated.
Bakar et al. presented an algorithm for finding minimum size
reducts which was based on rough set approach and a dedicated
decision-related binary integer programming (BIP) algorithm [8].
This algorithm transformed an equivalent class obtained from a
decision system into a BIP model. This work was linked to rough set
theory, data mining and non-monotonic reasoning.
Hu presented a new approach to construct a good ensemble of
classifiers using rough set theory and database operations [37].
Ensemble of classifiers was formulated precisely within the
framework of rough set theory and constructed very efficiently
by using set-oriented database operations. This method was used
to compute a set of reducts which included the entire indis-
pensable attribute required for the decision categories. For each
reduct, a reduct table was generated by removing those attributes
which were not in the reduct. Next a novel rule induction
algorithm was used to compute the maximal generalized rules for
each reduct table and a set of reduct classifiers were formed based
on the corresponding reducts.
Questier et al. described the use of rough set theory to construct
reducts in a supervised way for reducing the number of features in
an unsupervised clustering [83]. This feature selection method was
also applied in combination with clustering methods other than
the described hierarchical method.
Swiniarski and Skowron presented an application of rough set
method for feature selection in pattern recognition [93]. They
proposed a new feature selection method to the result of principle
component analysis (PCA) used for feature projection and
reduction. Finally rough set methods had shown ability to reduce
significantly the pattern dimensionality and had proven to be
viable data mining techniques as a front end of neural network
classifiers.
 K. Thangavel, A. Pethalakshmi / Applied Soft Computing 9 (2009) 1–12 3

Table 1
Feature selection for decision system based on the rough set theory approach

Authors Purpose Description

Golan and Ziarko [22]
Ruhe [85]
Duntsch and Gediga [18]
Lin [66]
Chan [11]
Starzik et al. [91]
Fujimori et al. [20]
Guan and Bell [29]
Griffin and Chen [23]
Khoo et al. [51]
Hu et al. [36]
Vinterbo and Ohrn [111]
Bakar et al. [7]
Bakar et al. [8]
Vesa et al. [109]
Yin et al. [118]
Phuong et al. [81]
Beynon [9]
Galvez et al. [21]
Diaz and Corchado [17]
Hu [37]
Li and Liu [62]
Haiying et al. [30]
Maji and Roy [69]
Questier et al. [83]
Zhang et al. [121]
Swiniarski and Skowron [93]
Wei [115]
Zhu and Wang [125]
Grzymala-Busse [25]
Muir et al. [74]
Chan [12]
De and Radhakrishna [16]
Zhang and Yao [122]
Yidong et al. [117]
Curry [15]
Feature selection Proposed rough set reduct methodology which used Datalogic/R
Feature selection Applied the rough set approach for analysis of software engineering data which
resulted from goal-oriented measurement
Feature selection and classification Developed a software system ‘GROBIAN’ which performed rough set data analysis
Feature selection Formulated and generalized rough set methodology relevant to granular computing
Classification Described a method for updating approximations incrementally which could
be used as an effective tool to deal with dynamic attribute generalizations
Feature selection Presented an expansion algorithm which allowed the generation of all reducts
in a much less time than the elimination method
Feature selection and classification Applied rough set theory to the discharge currents accompanied by tree
development and removed the data both horizontally and vertically
Feature selection and classification Presented an investigation of computational methods for rough analysis. The
significance measure allowed to reduce the computational cost of discovery
by the rough analysis method
Classification Introduced an approach which added a rough set extension to an existing
scripting language Tcl
Classification Proposed a novel approach for the classification and rule induction of
inconsistent information systems
Feature selection Proposed two reduct algorithms, one was based on attribute frequency in
discernibility matrix and another used similar idea and sampling techniques
for large data sets
Feature selection and classification Presented a problem of simplification method that ensured a lower bound of the
degree of approximation together with a genetic algorithm minimal cost
approximate hitting sets discovery
Classification Discussed the rough set theory on managing the uncertainty of decision in
the information systems. Solved the uncertain decision through measured the
lower and upper approximation, accuracy and quality of approximation
Feature selection Presented an algorithm based on rough set approach and a dedicated decision-
related binary integer programming (BIP) to find minimum size reducts
Feature selection and classification Described preprocessing, rule generation and classification validation
Feature selection Described a new approach for data filtering which effectively reduced granularity
of attribute measurement and improved the statistical significance of rules
Feature selection and classification Proposed an algorithm to infer rules from reducing and combining attributes
Feature selection Presented an investigation of the criteria for a b-reduct within variable precision
rough set (VPRS) model
Feature selection and classification Presented Conjuntos Aproximados con Incertidumbre (CAI) model, a new revision
of the variable precision rough set (VPRS) model with stronger rules which used
less number of attributes in their antecedents
Feature selection Proposed a feature subset selection algorithm based on a greedy strategy that tried
to maximize an information measure of entropy. It decreased the computation
effort for inducing a classifier
Feature selection and classification Presented a new approach which used rough set theory and database operations
to construct a good ensemble of classifiers
Feature selection Proposed the feature reduction algorithm and the conditional entropy to
compute the relevance of attributes
Feature selection and classification Presented the model and approach for hierarchical fault diagnosis for substation
based on rough set theory
Classification Applied the theory of soft sets to solve a decision-making problem which
used rough mathematics
Feature selection Described the use of rough set theory to construct reducts in a supervised
way for reducing the number of features in an unsupervised clustering
Feature selection Proposed a novel heuristic algorithm based on rough set theory to find out
the feature subset
Feature selection Presented an application of rough set method for feature selection in
pattern recognition
Feature selection Presented a new approach for the selection of attributes for the construction of
decision tree based on rough set theory. The results showed that the rough
set-based approach was a feasible way for the selection of nodes of decision tree
Feature selection Investigated some basic properties covering generalized rough sets. Constructed
a set of axioms to characterize the covering lower approximation operation
Classification Introduced a new algorithm called Modified Learning Examples Module
Version 2 (MLEM2). This algorithm produced minimum number of rules as
well as accuracy than LEM2, MODLEM Laplace and entropy-based algorithm
Feature selection Used binary decision diagram (BDD) for more efficient determination of
the discernibility function
Classification Presented a mechanism called LERS-M for learning production rules from
very large databases
Classification Presented a novel algorithm for clustering. This approach was useful to find
interesting user access patterns in weblog
Feature selection Proposed two new rough set theory-based feature selection techniques like average
support heuristics (ASH) and parameterized average support heuristic (PASH)
Classification Proposed two methods for rule inconsistency based on irregular decision tables
Feature selection and classification Developed sampling theory using both the original rough set model and its
probabilistic extension variable precision rough set (VPRS) model
4 K. Thangavel, A. Pethalakshmi / Applied Soft Computing 9 (2009) 1–12
Table 1 (Continued )
Authors Purpose
Kryszkiewicz and Cichon [55] Feature selection
Hu et al. [39] Feature selection
Chen et al. [13] Feature selection
Gu et al. [28] Feature selection and classification
Zeng and Zhan [120] Feature selection and classification
Tan [94] Classification
Tseng et al. [107] Classification
Wang [112] Classification
Thangavel et al. [96] Feature selection and classification
Pethalakshmi et al. [80] Feature selection
Thangavel et al. [100] Feature selection and classification
Thangavel et al. [104] Feature selection and classification
Grzymala-Busse introduced a new algorithm called Modified
Learning Examples Module Version 2 (MLEM2) [25]. It was
extended by LEM2 capabilities by inducing rules from data with
both symbolic and numerical attributes with missing attribute
values. It was compared with LEM2, MODLEM Laplace and
entropy-based algorithm. MLEM2 produces minimum number
of rules as well as accuracy than other algorithms.
Hu et al. proposed a new rough sets model and redefined the
core attributes and reducts based on relational algebra to take
advantages of the very efficient set-oriented database operations
[39]. Using this model and definitions, they presented two
algorithms for computing core and reduct. These algorithms were
also being applied in a real-life application with very large data
sets. These models were efficient and scalable compared with
traditional rough set models.
Chen et al. proposed a reasonable definition of parameteriza-
tion reduction of soft sets and compared it with the attribute
reduction in rough set theory [13]. By using this new definition of
parameterization reduction, they improved the application of a
soft set in a decision-making problem.
Thangavel et al. proposed an information system without any
decision attribute [96]. The proposed algorithm was useful when
the database contains only input information (condition attri-
butes) without decision (class attribute). K-Means algorithm was
applied to cluster the given information system for different values
of K. Decision table was formulated using this clustered data as the
decision variable. Then Quickreduct and VPRS algorithms were
applied for selecting features. Ultimately, Rule Algorithm was used
for obtaining optimum rules. The experiments were carried out on
data sets of UCI machine learning repository and the HIV data set to
analyze the performance study. It was observed that less number
of rules were produced when the VPRS reduct applied for K = 3
compared to Quickreduct.
In quickreduct algorithm [43], the vertical reduct was
possible, when only the unwanted attributes were eliminated.
Description
Proposed a number of algorithms to discover all generalized reducts (preserving
generalized decisions), all possible reducts (preserving upper approximations)
and certain reducts (preserving lower approximations)
Proposed a new rough sets model and redefined the core attributes and reducts
based on relational algebra to take advantages of the very efficient set-oriented
database operations
Proposed a reasonable definition of parameterization reduction of soft sets and
compared it with the attribute reduction in rough set theory
Presented the modified chimerge algorithm which eliminated the additional
information and the rules were extracted from rough set theory after continuous
values were discretized. This algorithm presented as a stopping criterion
compared to old chimerge algorithm
Proposed a new rough set-based feature selection. It reduced the number of features
and improved the classification accuracy
Presented an alternative streamlined approach wherein heuristic rules were derived
from a set of training data in the form of example alternatives ranked in order of
preference by the expert
Proposed rule-composing algorithm and rule-validation procedure based on
rough set theory
Proposed a tolerant rule induction algorithm for mining rules in noisy environments.
The result showed that the proposed algorithm was potentially more noise
insensible, more accurate and more cost effective than tree classification algorithm
Proposed an information system without any decision attribute. The unsupervised
technique was applied for clustering in this work
Proposed the attribute reduction for decision system contains multiple
decision attributes
Proposed Modified Quickreduct Algorithm for feature selection compared with
original Quickreduct and variable precision rough set (VPRS). The proposed
algorithm generated minimal reducts as well as less number of rules
Proposed an Accelerated Quickreduct algorithm to select the features from the decision
system followed by C4.5 algorithm for rule induction
As in the normalization process in the data base system, the size
of the information system could also be reduced horizontally by
eliminating the objects which are involved in the construction of
the lower approximation. Thangavel et al. [100] proposed
Modified Quickreduct algorithm for horizontal and vertical
reduction and discussed the performance study of various
reduct algorithms such as Quickreduct and Variable Precision
Rough Set (VPRS) for constructing the efficient rules. The
experiments were carried out on data sets of UCI machine
learning repository and the human immune deficiency virus
(HIV) data set to analyze the performance study. Generally, in
rule generation for taking decision from the information system,
the reduct plays a vital role. The reduct algorithm that generated
the least number of rules was considered an efficient one. It was
observed that the Modified Quickreduct generated minimal
reducts. The rule induction was performed by using minimal
reduct set generated by Modified Quickreduct and it was
identified that less number of rules were produced when
compared with the rules generated by the reducts that used
Quickreduct and Variable Precision Rough Set.
Thangavel and Pethalakshmi proposed an Accelerated Quick-
reduct algorithm to select the features from the decision system
[104] followed by the C4.5 for rule induction. A comparative study
was also performed with the proposed and original Quickreduct
algorithms. The experiments were carried out on public domain
datasets available in UCI machine learning repository to analyze
the performance study. It was evident from the performance
analysis that the Accelerated Quickreduct produced the minimal
reduct and the rules induced through C4.5 algorithm revealed that
the Acclerated Quickreduct was performing well.
3. Hybrid approach
This section describes the hybrid approaches of various
techniques for feature selection and classification.
 K. Thangavel, A. Pethalakshmi / Applied Soft Computing 9 (2009) 1–12 5

3.1. Feature selection and classification using rough and fuzzy sets
The use of fuzzy set theory is one way of capturing the
vagueness present in the real world, which would otherwise be
difficult using conventional set theory. There are many useful
introductory resources regarding fuzzy set theory [14,78]. In
classical set theory, the elements could belong fully (i.e. have a
membership of 1) or not al all (a membership of 0). Fuzzy set
theory relaxes this restriction by allowing memberships to take
values anywhere in the range [0, 1]. A fuzzy set can be defined as a
set of ordered pairs A = {x, mA(x) j x 2 U}. The function mA(x) is
called the membership function for A, mapping each element of the
universe U to a membership degree in the range [0, 1]. The universe
may be discrete or continuous. Any fuzzy set containing at least
one element with a membership degree of 1 is called normal.
The various applications for feature selection and classification
using rough and fuzzy sets are summarized in Table 2.
Rough set theory allows for obtaining a linguistic description of
the function whereas the fuzzy logic theory allows to generate
numerical values of the function starting from its linguistic
description.
Alonge et al. studied the problem of identification of a non-
linear vectorial function based on rough set theory and fuzzy logic
[3]. The covering index gave good results for determining
approximate rules from rough ones.
Rough set theory is reliant upon a crisp dataset that is important
information may be lost as a result of quantization. To avoid this
information loss, Jensen and Shen reviewed semantics-preserving
dimensionality reduction technique that preserved the underlying
semantics of the data using crisp and fuzzy-rough set-based
methodologies [46]. The result showed that the conventional
rough set methods were unable to deal with real-valued attributes
effectively. This prompted research into the use of fuzzy-rough sets
for feature selection.
Tsai et al. proposed a new fuzzification technique called
modified minimization entropy principle algorithm (MMEPA) to
construct membership functions of fuzzy sets of linguistic
variables [106]. This technique was combined with variable
precision rough set (VPRS) model to form an entropy-based
fuzzy-rough classification approach.
Hong et al. proposed a method which combined the variable
precision rough set (VPRS) model and fuzzy set theory for solving
the problem of producing a set of fuzzy certain and fuzzy possible
rules from quantitative data with a predefined tolerance degree of
uncertainty and misclassification [34]. Rough set theory trans-
formed each quantitative value into a fuzzy set of linguistic terms
that used membership functions and then calculated fuzzy b-
lower and fuzzy b-upper approximations. Then certain and
possible rules were generated based on these fuzzy approxima-
tions. These rules were used to classify unknown objects.
Thangavel and Pethalakshmi proposed an Improved Quickre-
duct algorithm based on rough and fuzzy sets to select the features
from the information system [97]. The given information system
was normalized in the proposed algorithm and the computational
procedure for computing degree of dependency of each attribute
had been modified suitably. It was evident from the performance
analysis that the Improved Quickreduct produced the minimal
reduct from the data set containing large number of attributes with
large data.
3.2. Rough sets and neural network
A neural network is a technique that seeks to build an
intelligent system using models that simulate the working
network of the neurons in the human brain [35]. A neuron is
made up of several protrusions called dendrites and long-branch
called the axon. A neuron is joined to other neurons through the
dendrites. The dendrites of different neurons meet to form
synapses, the areas where message pass. The neurons receive
the impulses via the synapses. If the total of the impulses received
exceeds a certain threshold value, then the neuron sends an
impulse down the axon where the axon is connected to other
neurons through more synapses. The synapses may be excitatory
or inhibitory in nature. An excitatory synapse adds to the total of
the impulses reaching the neuron, whereas an inhibitory neuron
reduces the total of the impulses reaching the neuron. In a global
sense, a neuron receives a set of input pulses and sends out another
pulse that is a function of the input pulses.
The methods for feature selection and classification using rough
set and neural network are summarized in Table 3.
Li and Wang presented a hybrid approach of integrating rough
sets and neural networks to mine classification rules from large
data sets [63]. They also proposed a new algorithm for finding a
reduct and a new algorithm for rule generation from a decision
table based on a binary discernibility matrix. The reduct was
obtained using rough set theory and the neural network was
applied to delete noisy data. Again applied the rough set theory to
obtain rules or patterns. This hybrid approach generated a more
concise and accurate rules than traditional neural network based
approach and rough set-based approach.
Wang et al. described the relevance feedback. It was the most
important image retrieval system. They proposed a new relevance
feedback algorithm ARFRS [113]. This algorithm was composed of
two parts (i) preprocessing part and (ii) classification part. In
preprocessing part rough set theory was used to induce the
classification rules and reduct set of the image features. Neural
network technique was applied for classification to train the
retrieval return images based on the classification rules and reduct
set. This algorithm made full aspects of the advantage of rough set
theory and neural network ensembles to represent the human
perception objectively and improve the retrieval performance.
3.3. Metaheuristic algorithms with rough set theory
A metaheuristic is a set of algorithmic concepts that can be used
to define heuristic methods applicable to a wide set of different
problems. In other words, a metaheuristic can be seen as a general-
purpose heuristic method designed to guide an underlying

Table 2
An overview of rough and fuzzy sets

Authors Description

Alonge et al. [3] Studied the problem of identification of a non-linear vectorial function based on rough set theory and fuzzy logic
Jensen and Shen [46] Proposed a reduct algorithm based on rough and fuzzy sets
Tsai et al. [106] Proposed a new fuzzification technique to construct membership functions of fuzzy sets of linguistic variables
Hong et al. [34] Proposed a new method which combines the variable precision rough set (VPRS) model and fuzzy set theory
from quantitative data with a predefined tolerance degree of uncertainty and misclassification
Thangavel and Pethalakshmi [97] Proposed Improved Quickreduct Algorithm based on rough and fuzzy sets to select the features from
the information system
6 K. Thangavel, A. Pethalakshmi / Applied Soft Computing 9 (2009) 1–12
Table 3
An overview of rough sets and neural network
Authors Description
Mak and Munakata [70] Reviewed and compared the rule extraction capabilities of rough sets with neural networks and ID3
Li and Wang [63] Presented a hybrid approach of integrating rough sets and neural networks to mine classification rules from large data sets
Wang et al. [113] Described the relevance feedback for image retrieval system and proposed a new relevance feedback algorithm ARFRS
problem specific heuristic toward promising regions of the search
space containing high quality solutions. A metaheuristic therefore
a general algorithmic framework, which can be applied to different
optimization problems with relatively few modifications to make
them, adapted to a specific problem. The use of metaheuristics has
significantly increased the ability of finding very high-quality
solutions to hard, practically relevant combinatorial optimization
problems in a reasonable time. This is particularly true for large
and poorly understood problems. The metaheuristic algorithms for
feature selection and classification are summarized in Table 4.
Jagielska et al. compared the rules between neural network and
genetic algorithm, fuzzy rule induction systems and rough set
theory with C4.5 inductive algorithm [42]. Among these genetic
algorithm/fuzzy system provided the rule sets with higher
accuracy and comprehensibility.
Mitra et al. presented a methodology for integrating four soft
computing tools such as artificial neural network (ANN), fuzzy sets
(FS), genetic algorithms (GA) and rough sets (RS) for designing a
knowledge based network for pattern classification and rule
generation [72]. The proposed method involves several fuzzy MLP
modules, each encoding the rough set rules for a particular class.
These knowledge-based modules were refined using a genetic
algorithm. The role of fuzzy sets were to handle uncertainty with
input–output membership functions, while rough sets exploited
the granularity in information to obtain dependency rules used for
knowledge encoding in ANN. The final rules represented refined
knowledge extracted from the trained network and are more
meaningful and compact.
Jensen and Shen proposed a feature selection mechanism based
on ant colony optimization (ACO) [44]. They also presented an
entropy-based modification of the original rough set-based
approach. This algorithm was applied to the problem of finding
minimal rough set reducts.
Table 4
An overview of metaheuristic algorithms
Authors Description
Jagielska et al. [42] Compared various data mining techniques for rule
identification
Mitra et al. [72] Presented a methodology for integrating four soft
computing tools such as artificial neural network
(ANN), fuzzy sets (FS), genetic
algorithms (GA) and rough sets (RS) for designing
a knowledge based network for pattern
classification and rule generation
Jensen and Shen [44] Proposed a new feature selection mechanism based
on ant colony optimization (ACO)
Jaganathan et al. [40] Presented a system that combined both Quickreduct
algorithm for data preprocessing, Ant Miner and
C4.5 for classification. Ant Miner
produced a higher accuracy rate and fewer rules
than decision tree induction (C4.5)
Jaganathan et al. [41] Presented a system that combined both Improved
Quickreduct algorithm for feature selection and
Ant Miner for classification. It was demonstrated
that Ant Miner with improved Quickreduct
produced a higher accuracy rate and fewer rules
than the original Ant miner algorithm
Thangavel et al. [103] Presented a survey of ant colony algorithms in
diverse combinational optimization problems
Jaganathan et al. presented a system that combined both
Improved Quickreduct algorithm for feature selection and Ant
miner for classification [41]. The proposed system used the Ant
miner software and was tested on standard problems. It was
demonstrated that Ant miner with improved Quickreduct pro-
duced a higher accuracy rate and fewer rules than the original Ant
miner algorithm. The performance of the ant miner was increased
when it was used with Improved Quickreduct.
Thangavel et al. presented a survey of ant colony algorithms in
diverse combinational optimization problems [103]. It was
intended to summarize the methods of ant colony system used
in various types of applications. In particular, routing, assignment,
scheduling, subset, machine learning and network routing
problems.
4. Decision system with missing data
In real-life data, some of the attribute values are frequently
missing, imperfect, and incomplete. There are two main reasons for
missing attribute values: either they are ‘lost ‘or they ‘do not care’
conditions. That is, the original values were not recorded at all
since they were irrelevant, and the decision to which concept a
case belongs was taken without that information. Missing data is a
common problem in statistical analysis. In particular, missing
values in a data set can affect the performance of a classifier
constructed using such a data set as a training sample. Rates of less
than 1% missing data are generally considered trivial, 1–5%
manageable. However, 5–15% must be handled by complicated
methods, and more than 15% may severely impact any kind of
interpretation [82].
The techniques for feature selection and classification using
rough set for decision system with missing attribute values are
summarized in Table 5.
Grzymala-Busse and Hu presented and compared nine different
approaches to missing attribute values [24]. Ten input data files
were used to investigate the performance of the nine methods to
deal with missing attribute values. For testing both naı̈ve and new
classification techniques of learning from examples based on
rough sets (LERS) were used. The quality criterion of the average
error rate was achieved by 10-fold cross-validation.
Hong et al. dealt with the problem of producing a certain and
possible rules from incomplete datasets based on rough sets [33].
He proposed a new learning algorithm which was simultaneously
derived rules from incomplete datasets and estimated the missing
values in the learning process. Unknown values were first assumed
by any possible values and were gradually refined according to the
incomplete lower and upper approximations derived from the
given training examples. The examples and the approximations
interacted with each other to derive certain and possible rules and
estimated appropriate unknown values.
Latkowski and Mikolajczyk presented a new approach to handle
incomplete information and classifier complexity reduction [59].
They described a method called D3RJ that performed data
decomposition and decision rule joining. The original incomplete
decision table was decomposed into sub-tables without missing
values. Next the methods for induction of decision rules were
applied to these sets. Finally an algorithm for decision rule joining
 K. Thangavel, A. Pethalakshmi / Applied Soft Computing 9 (2009) 1–12 7

Table 5
An overview of the techniques for decision system with missing attribute values

Authors Description

Slowinski and Stefanowski [90]
Jiye and Zongben [48]
Grzymala-Busse and Hu [24]
Hong et al. [33]
Grzymala-Busse [26]
Grzymala-Buss and Siddhaye [27]
Latkowski and Mikolajczyk [59]
Leung et al. [60]
Thangavel et al. [101]
Investigated the case of incomplete information systems and presented a generalization of the rough sets approach which
dealt with missing and imprecise descriptors
Described an uncertainty measures of roughness of knowledge and rough sets by introducing rough entropy in incomplete
information systems
Presented and compared nine different approaches to missing attribute values
Proposed a new learning algorithm which was simultaneously derived rules from incomplete datasets and estimated
the missing values in the learning process
Presented a data set in the form of incompletely specified decision table and also introduced three definitions of lower
and upper approximations
Presented the decision tables as incomplete and also proposed Modified Learning Examples from Module Version 2
(MLEM2) rule induction algorithm
Presented a new approach called D3RJ to handle incomplete information and classifier complexity reduction
Proposed the similarity classes in incomplete information systems and two kinds of approximations are formed for the
mining of certain and association rules in incomplete decision tables
Described four different methods for dealing with missing attribute values and proposed a Revised Quickreduct algorithm
for dimensionality reduction

was used to obtain the final rule set from partial rule sets. It
produced small set of rules and better classification accuracy than
the classic decision rule induction methods.
Thangavel et al. discussed four different methods such as
indiscernibility, mean, median and mode for dealing with missing
attribute values and proposed a revised Quickreduct algorithm for
dimensionality reduction [101]. A comparative study was also
performed with revised and original Quickreduct algorithms based
on the four different methods. The public domain datasets available
in UCI machine learning repository with missing attribute values are
used. It was found that the revised Quickreduct algorithm out-
performed the original Quickreduct algorithm when there were
large numbers of patterns with more number of attributes.
5. Applications
Rough set theory has been applied successfully almost in all the
areas. One of the major limitations of the traditional rough set
model in the real applications is the inefficiency in the computa-
tion of core attributes and the generation of reducts. In order to
improve the efficiency of computing core attributes and reducts,
many novel approaches have been developed [39,53]. This section
shows the various applications for feature selection and classifica-
tion. These applications are summarized in Table 6.
Kusiak developed an autonomous decision-making based on
rough set theory. This approach was tested on a medical dataset for
patients with lung abnormalities referred to as solitary pulmonary
nodules (SPNs) [57]. Two independent algorithms were developed
either generate an accurate diagnosis or make no decision. The
feature extraction algorithm generated reducts. These reducts
were used for predicting objects outcome with the primary and
confirmation algorithms. The dataset contained 18 features. Final
results showed that the classification quality by the combined
primary and confirmation algorithms was 91.3% (i.e.) 91.3% of all
patients tested had been correctly diagnosed and the diagnostic
accuracy was 100%. Finally it showed that the reduced number of
attributes should lower testing costs.
For feature selection, rough set theory has been used as such a
dataset preprocessor with much success, but current methods are
inadequate for finding minimal reductions, the smallest sets of
features possible. To alleviate this difficulty, a feature selection
technique that employs a hybrid variant of rough sets, fuzzy-rough
sets, has been developed recently and has been shown to be
effective. Jensen and Shen described the shortcomings of tradi-
tional rough set attribute reduction when applied to datasets with
real-valued attribute and they proposed a new method based on
fuzzy-rough sets [43]. This approach incorporated the information
usually lost in crisp discretization by utilizing the generated fuzzy-
rough sets to provide a more informed technique. This approach
performed better than traditional rough set attribute reduction on
the whole, in terms of both data reduction and reduction quality.
They evaluated using the new fuzzy rough set attribute reduction
(RSAR) approach for the water treatment plant database. The
dataset contains 38 condition attributes and a decision attribute.
The database results were compared with crisp RSAR approach.
The experimental analysis showed that fuzzy RSAR approach
selected fewer attributes as well as lower classification error over
the entire tolerance range than crisp RSAR.
Midelfart et al. presented a general rough set approach for
classification of tumor samples analyzed with micro-arrays [71].
This approach was tested on a data set of gastric tumors and they
developed classifiers for six clinical parameters. One major obstacle
in training classifiers from micro-array data is that the number of
objects is much smaller than that the number of attributes. So they
introduced a feature selection method based on boot strapping for
selecting genes that discriminate significantly between the classes.
This method was examined with the Rosetta system [105]. Their
performance was compared to that of linear and quadratic
discrimination analysis. The classifiers were also biologically
validated. One of the best qualifiers was selected for each clinical
parameter. And the connection between the genes used in these
classifiers and the parameters were compared to establish knowl-
edge in the biomedical literature. The result showed that it was
possible to develop classifiers with the small number of tumor
samples and this method was well suited for this task.
Shen and Chouchoulas proposed a highly modular framework for
data-driven fuzzy rule set induction incorporating a dimensionality
reduction step based on rough set theory [87]. This removed
redundant and information-poor attributes from the data, thereby
significantly increasing the speed of the induction algorithm which
was employed to generalize historic data into fuzzy association
rules. They demonstrated this approach using Water Treatment
Plant database. The dataset contains 38 attributes. Rough set
attribute reduction (RSAR) was used to reduce the dataset. Then rule
induction algorithm (RIA) was used to classify the data. Then RIA was
compared with C4.5 algorithm. RIA produced classification accuracy
of 96.5% (two-class) and C4.5 produced 96.8% (two-class). Hence
C4.5 produced better accuracy than RIA.
Shen and Chouchoulas presented an approach that integrated a
potentially powerful fuzzy rule induction algorithm with a rough
set-assisted feature reduction method [88]. They proposed Quick-
reduct II algorithm for feature selection. They demonstrated this
8 K. Thangavel, A. Pethalakshmi / Applied Soft Computing 9 (2009) 1–12
Table 6
An overview of applications
Authors
An et al. [5]
Komorowski and Ohrn [53]
Nakayama et al. [75]
Kusiak [56]
Kusiak et al. [57]
Kusiak et al. [58]
Breault [10]
Swiniarski [92]
Abidi et al. [1]
Li and Jiang [65]
Enrique [19]
Zhong and Skowron [123]
Zhong et al. [124]
Hu and Cercone [38]
Vesa et al. [110]
Abidi and Hoe [2]
Midelfart et al. [71]
Jensen and Shen [43]
Shen and Chouchoulas [87]
Shen and Chouchoulas [88]
Tay and Shen [95]
Shen and Jensen [89]
Li et al. [64]
Krishnaswamy et al. [54]
Warren et al. [114]
Applications
Water demand
Cardiac patients
Diabetic database
Industrial data
Solitary pulmonary nodules (SPNs)
SPN and engineering datasets
Pima Indian Diabetic Database
Face recognition
New thyroid gland (NTG),
Wisconsin breast cancer (WBC)
Multisensor datum
Ergonomic diagnosis
Slope-collapse database
Slope-collapse and gastric cancer
Market database
Test network
New thyroid gland (NTG),
Wisconsin breast cancer (WBC)
Gastric tumors
Water Treatment Plant database
Water Treatment Plant database
Water Treatment Plant database
Multi-cylinder diesel engine
Water Treatment Plant database
Text categorization
Datamining and high performance
computing
Solid waste management (SWM)
Methods
KDD-R
ROSETTA
Rough analysis
Datamining approach
Primary and confirmation
based on rough analysis
Rough analysis
ROSETTA
PCA, LVQ
Hybrid rule extraction (data clustering,
data discretization, attribute selection,
symbolic rule discovery)
Mixed VCR, mixed fuzzy inference and
pure fuzzy inference
Rough analysis technique
Rough sets heuristics (RSH) and rough
sets with boolean reasoning (RSBR)
Rough analysis and greedy heuristics
DB Deci, C4.5
Fuzzy C-means clustering (FCM),
rough sets
Hybrid rule extraction (data
clustering, data discretization,
symbolic rule discovery)
Dynamic reduct, IR classifier and
genetic algorithm
Fuzzy-rough feature selection (FRFS)
Fuzzy rule induction algorithm
(RIA), C4.5
Quickreduct II, RIA, C4.5
Rough analysis
Fuzzy-rough feature selection
(FRFS), conventional entropy, PCA,
random-based
Case-based reasoning (CBR)
Rough analysis technique
Modified RS1
Results
Best error rate was 6.67% and average
error rate prediction was 10.27%
Described decreasing the degrees of
precision p, increases the accuracy.
If p = 1.0, accuracy = 88.3% and if p
= 0.5, accuracy = 95.9%
Fitting rate 81.7% or 84.9% (fuzzy inference)
was high for supporting diagnosis of
Macroangiopathy
Extracted the knowledge for prediction and
prevention manufacturing faults in wafers
Classification quality was 91.3% with
diagnostic accuracy 100%
Classification quality of SPN was 91.3%
with diagnostic accuracy 100%. For
Engineering dataset the classification
quality was 96.8%
The mean accuracy was 73.8% with 95% CI
where the previous researches produced
71.5% mean accuracy with 63.3% CI
Produced the classification accuracy
of 97.3% for the test set
Overall predictive accuracy of NTG
was 86% and WBC was 87%
In an average, mixed fuzzy inference
method produced the highest classification
accuracy of 88% than other two methods
Described a rough analysis as a
technique to direct the knowledge
discovery process from data applied
to ergonomic diagnosis
Among 24 attributes, 9 were selected
Among 24 attributes, 8 were selected
for slope-colapsed and among 50
attributes 19 were selected (class =
cause of death) and 17 were selected
(class = post-operative complication)
Average prediction accuracy of DB Deci
was 73.8% whereas C4.5 was 67.3%
In an average 95% of data are classified
successfully. Rough sets generated 193
rules whereas FCM produced 25 rules
Produced less than 30 rules for NTG
and WBC
This method was examined with ROSETTA
system. Most of the classifiers had a very
good accuracy and high AUC value
Fuzzy rough method selected fewer
attributes and lower classification error
than crisp rough set
RIA produced classification accuracy of
96.5% and C4.5 produced 96.8%. C4.5
was better than RIA
RIA produced classification accuracy of
97.3% while C4.5 produced 96.8%. RIA
was better than C4.5
Distinguished the fault types or inspect
the dynamic characteristic of the machinery
FRFS produced better accuracy (training
accuracy 83.3% and testing accuracy
83.9%) than other methods
Proved that the performance of document
reduction depends on the threshold ‘thr’.
In an average if thr = 0.96, 43.6% of
documents were reduced
Measured the error rate of San Diego
Super Computing (SDSC) ‘95 and ‘96.
SDSC 95 was better than SDSC 96 on
the average mean error
Enabled to handle larger datasets as
previous RS1
 K. Thangavel, A. Pethalakshmi / Applied Soft Computing 9 (2009) 1–12 9

Table 6 (Continued )
Authors Applications Methods Results

Zaluski et al. [119]
An et al. [6]
Tsumoto [108]
Salem et al. [86]
Jensen and Shen [45]
Hassanien [31]
Hassanien and Ali [32]
Jensen and Shen [47]
Li and Cercone [61]
Wilk et al. [116]
Pethalakshmi and Thangavel [79]
Thangavel et al. [98]
Thangavel et al. [99]
Karnan et al. [50]
Thangavel et al. [102]
Breast cancer LEM2, C4.5, CN2, Assistant86 Accuracy of original dataset was 70.63%
and standard deviation was 8.4%. For
reduced dataset, the accuracy was 71.5%
and standard deviation was 6.06%
Web page classification ELEM2 Rough set feature selection could improve
the predictive performance when the
original feature set for representing web
pages is large
Head-ache Rough analysis, PRIMEROSE4, Proposed algorithm produced the average
C4.5, AQ15 accuracy of 95%, PRIMEROSE4 88.3%, C4.5
75.8% and AQ15 79.3%
Thrombosis database Rough analysis Out of 60, 16 attributes were selected
Bookmark classification and Crisp rough set attribute reduction, In bookmark classification, the classification
website classification fuzzy-rough feature selection, accuracy of BIM was 94.3% and VSM was
vector space model (VSM), boolean 98.1%. In website classification using crisp
inexact model (BIM) method, BIM classifier produced average
precision and average error rates were
23.3%, 88.3%, respectively. In VSM
classifier 45.2%, 49.7%, respectively. In
fuzzy method, BIM produced the average
precision and average error rates were
16.7%, 80%, respectively and VSM produced
74.9%, 35.9%, respectively. In fuzzy RSAR
approach, VSM classifier produced better
average precision and
minimum average error than BIM method
Breast cancer Rough analysis Rough set was a useful tool for inductive
learning and valuable aid for building
expert systems
Mammogram dataset Rough analysis, decision trees, Classification accuracy of rough set was
neural networks 93.1%, decision Trees 83% and neural
networks 90.34%
Complex Systems monitoring antFRFS, FRFS, PCA, entropy-based antFRFS outperformed than other methods
Geriatric dataset ROSETTA Generated rules with support = 30% and
confidence = 80%
Abdominal pain LEM2, explore Explore produced 83.99% classification
accuracy whereas LEM2 produced 74.07%
HIV LEM2, MLEM2 Less number of rules were generated in
MLEM2 rather than LEM2. MLEM2
generated rules from raw data with
numerical attributes, without any
prior discretization
MIAS Decision relative discernibility-based Among the five, the decision relative
reduction, heuristic approach, Hu’s discernibility-based reduction and heuristic
algorithm, Quickreduct, VPRS algorithms produced minimal features
UCI medical dataset Improved Quickreduct, Improved Quickreduct produced minimal
Quickreduct, C4.5 reduct as well as minimal rule from the
dataset containing large number of attributes
MIAS Backpropagation neural network, BPN-ACO classifier performed better than
backpropagation classifier hybrid other classifiers with higher percentage
with ant colony. Optimization of classification rate
MIAS Genetic algorithm, ant colony A higher Az indicated better classification
optimization, rough set-based performance because a larger value of
reduction algorithms, BPN, true positive (TP) was achieved at each
Jack-Knife Method, Round value of false positive (FP)
Robin Method

approach using real world problems like water treatment plant data
and algae population estimation data. Water treatment data consists
of 38 attributes and in algae population, seven alga species was
processed separately. The reduct set was computed using the
Quickreduct II algorithm. Then apply the rule induction algorithm
such as fuzzy rule induction algorithm (RIA) and C4.5. In water
treatment plant data, RIA produced the classification accuracy of
97.3% while C4.5 produced 96.8%. In algae data, RIA produced
minimal error rate than C4.5. Hence RIA is better than C4.5.
Shen and Jensen proposed a feature selection technique that
employed a hybrid variant of rough sets, fuzzy-rough sets to avoid
the information loss [89]. This proposed method was applied to
Water Treatment Plant database. The final results showed that the
proposed method was more powerful than conventional entropy,
principle component analysis (PCA) and random-based method.
The Water Treatment Plant database contains 38 condition
attributes and a decision attribute. The proposed method was
compared with entropy-based feature selection (number of
classes = 2), PCA and random-based methods. The training
accuracy was 83.3% and testing accuracy was 83.9% in the
proposed method. Entropy-based algorithm produced the training
and testing accuracy as 80.7% and 83.9%, respectively. In random-
based method training accuracy was 66.4% and testing accuracy
was 68.1%. In PCA the training accuracy was 76.7% and testing
accuracy was 70.3%. So rule complexity was greatly reduced in the
proposed method. Similarly in three-class dataset, the training
accuracy was 70% and testing accuracy was 71.8% in the proposed
method. Entropy-based algorithm produced the training and
testing accuracy as 70% and 72.5%, respectively. In random-based
method training accuracy was 55.7% and testing accuracy was
10 K. Thangavel, A. Pethalakshmi / Applied Soft Computing 9 (2009) 1–12
54.3%. In PCA the training accuracy was 67.7% and testing accuracy
was 70.2%. On the whole proposed method produced highest
classification than PCA and random-based methods over the
tolerance range.
Hassanien and Ali presented an enhanced rough set approach
for attribute reduction and generating classification rules from
digital mammogram datasets [32]. They compared rough set
results with decision trees and neural networks. The classification
accuracy of rough sets was 93.1%, decision trees provided 83% and
neural networks provided 90.34%. The final result showed that
rough sets outperformed compared with decision trees and neural
network models.
Fuzzy rough feature selection (FRFS) was not able to find the
optimal subsets regularly. So, Jensen and Shen proposed a new
feature selection mechanism based on ant colony optimization
(ACO) [47]. The method was applied to the problem of finding
optimal feature subsets in the fuzzy-rough data reduction
process. To demonstrate this method using complex systems
monitoring and experimentally compared with the original
fuzzy-rough method (FRFS), an entropy-based feature selector,
and principal component analysis (PCA). For both the two-class
and three-class datasets, FRFS and antFRFS select at least three
fewer features than the entropy-based method. However, the
entropy-based method outperformed the other two feature
selectors with the resulting C4.5 classification accuracies. This
was probably due to the fact that C4.5 used exactly the same
entropy measure in generating decision trees. In this case, the
entropy-based measure would favour those attributes that
would be the most influential in the decision tree generation
process. In the case of PCA, both antFRFS and FRFS significantly
outperform PCA on the two-class dataset. The resulting error for
PCA was 19.7% for the training data and 26.7% for the test data.
For antFRFS the errors were 6.5% (training) and 22.1% (testing),
and for FRFS 10.8% (training) and 25.2% (testing). In the three-
class dataset experimentation, both fuzzy-rough methods
produced much lower classification errors than PCA for the
training data. For the test data, the performance was about the
same, with PCA producing a slightly lower error than antFRFS on
the whole. As a result, antFRFS produced minimal reduct than
other algorithms.
Segmentation and classification of objects is an important area
of research and of practical applications in a variety of fields,
including pattern recognition and artificial intelligence, medicine,
statistics and vision analysis. Thangavel et al. presented the
textural analysis method spatial gray-level dependency method
(SGLDM) was used to extract the features from the segmented
image [102]. From the extracted features set, the optimum
features were selected using genetic algorithm, ant colony
optimization and rough set-based reduction algorithms such as
decision relative discernibility-based reduction, Heuristic
approach, Hu’s algorithm, Quickreduct (QR), and variable preci-
sion rough set (VPRS). The selected textural features were input to
a three-layer backpropagation neural network (BPN) classifier to
classify the microcalcifications into benign or malignant. The BPN
classifier was validated using Jack-Knife Method, Round Robin
Method and Ten-Fold Validation method. To evaluate the
classification performances of the selected textural features, a
receiver operating characteristics (ROC) analysis was performed.
The area under the ROC curve Az was used as a measure of the
classification performance. A higher Az indicated better classifica-
tion performance because a larger value of true positive (TP) was
achieved at each value of false positive (FP). The above algorithms
and the techniques were tested on 161 pairs of digitized
mammograms from Mammography Image Analysis Society
(MIAS) database.
6. Conclusion
In this survey we have reviewed research on the application of
techniques from rough set, fuzzy set, neural network and
metaheuristic to select more dominating features in order to
perform the classification tasks optimally in the data mining
process. Several researchers have contributed variety of algorithms
for computing the reduct sets by considering different cases like
inconsistency, missing attribute values, multiple decision attri-
butes of the decision system and information system with no
decision attributes. The relationship and performance analysis of
the reduct algorithms have been discussed in connection with the
classification for various applications. The incremental algorithms
are needed for feature selection in the real-life applications. This is
the direction for further research.
Acknowledgements
The authors would like to thank all the anonymous reviewers
for their comments to improve the quality of the paper.
References
[1] S.S.R. Abidi, K.M. Hoe, A. Goh, Analyzing Data Clusters: A Rough Set Approach To
Extract Cluster-Defining Symbolic Rules, LNCS 2189, in: Proceedings of the IDA,
2001.
[2] S.S.R. Abidi, K.M. Hoe, Symbolic exposition of medical data-sets: a data mining
workbench to inductively derive data-defining symbolic rules, in: Proceedings of
the IEEE Symposium on Computer Based Medical Systems, 2002.
[3] F. Alonge, F. D’ Ippolito, G. Pizzuto, F.M. Raimondi, Identification of non linear
vectorial functions by means of rough set and fuzzy logic theories, in: Proceed-
ings of the IEEE International Conference on Control Applications, 1998.
[4] J.J. Alpigini, J.F. Peters, J. Skowron, N. Zhong, Rough sets and current trends in
computing, in: Proceedings of the 3rd International Conference, USA, 2002.
[5] A. An, C. Chan, N. Shan, N. Cercone, W. Ziarko, Applying knowledge discovery to
predict water-supply consumption, Expert IEEE 12 (4) (1997) 72–78.
[6] A. An, Y. Huang, X. Huang, N. Cercone, Feature Selection with Rough Sets for Web
Page Classification Transactions on Rough Sets, SpringerLink Publishers, 2004,
pp. 1–13.
[7] A.A. Bakar, Md.N. Sulaiman, M. Othman, M.H. Selamat, On managing the uncer-
tain decision: a rough set approach, in: Proceedings of the National Conference
on Management Science/Operations Research, 2000.
[8] A.A. Bakar, Md.N. Sulaiman, M. Othman, M.H. Selamat, Finding minimal reduct
with binary integer programming in data mining, in: Proceedings of the TENCON,
2000, pp. 141–149.
[9] M. Beynon, Reducts within the variable precision rough sets model: a further
investigation, European Journal of Operational Research 134 (2001) 592–605.
[10] J.L. Breault, Data mining diabetic databases: are rough sets a useful addition?
Proceedings of the Computing Science and Statistics 33 (2001) 12001.
[11] C.-C. Chan, A rough set approach to attribute generalization in data mining,
Information Sciences 107 (1998) 169–176.
[12] C.-C. Chan, Learning rules from very large databases using rough multisets,
Transactions on Rough Sets 1 LNCS 3100 (2004) 55–77.
[13] D. Chen, E.C.C. Tsang, D.S. Yeung, X. Wang, The parameterization reduction of
soft sets and its applications, International Journal on Computers and Mathe-
matics 49 (2005) 757–763.
[14] E. Cox, The Fuzzy Systems Handbook: A Practioner’s Guide to Building Using and
Maintaining Fuzzy Systems, Academic Press Inc., 1999, , ISBN: 0-121-94270-8.
[15] B. Curry, Sampling aspects of rough set theory, Computational Management
Science 1 (2004) 151–178.
[16] S.K. De, P. Radha Krishna, Clustering web transactions using rough approxima-
tion, Fuzzy Sets and Systems 148 (2004) 131–138.
[17] F. Diaz, J.M. Corchado, The selection of relevant features and rough sets, in:
Proceedings of 9th International Conference España, 2001.
[18] I. Duntsch, G. Gediga, The rough set engine GROBIAN, in: Proceedings of the
IMACS World Congress, vol. 4, 1997, pp. 613–619.
[19] A. Enrique, H. Aquiles, T. Nuria, G. Rosario, Rough analysis applied to ergonomics:
a first approximation, in: Proceedings of Computer-Aided Ergonomics and
Safety, 2001.
[20] S. Fujimori, K. Mikami, T. Kajiya, T. Inoue, Study on discharge currents with rough
set theory, in: Proceedings of IEEE International Symposium on Electrical
Insulation, 1998, pp. 432–435.
[21] J.F. Galvez, F. Diaz, P. Carrion, A. Garcia, An application for knowledge discovery
based on a revision of VPRS model, in: Rough Sets and Current Trends in
Computing 2000, LNAI 2005, 2001, 296–303.
[22] R.H. Golan, W. Ziarko, A methodology for stock market analysis utilizing rough
set theory, Proceedings of the IEEE/IAFE (1995) 32–40.
 K. Thangavel, A. Pethalakshmi / Applied Soft Computing 9 (2009) 1–12
[23] G. Griffin, Z. Chen, Rough set extension of Tcl for data mining, Knowledge-Based
Systems 11 (1998) 249–253.
[24] J.W. Grzymala-Busse, M. Hu, A comparison of several approaches to missing
attribute values in data mining, in: W. Ziarko, Y. Yao (Eds.), RSCTS 2000, LNAI
2005, 2001, pp. 378–385.
[25] J.W. Grzymala-Busse, MLEM2-discretization during rule induction, Procedings
of IIPWM (2003).
[26] J.W. Grzymala-Busse, Rough set strategies to data with missing attribute values,
in: Proceedings of the Workshop on Foundations and New Directions in Data
Mining, Associated with the 3rd IEEE International Conference on Data Mining,
2003, pp. 56–63.
[27] J.W. Grzymala-Busse, S. Siddhaye, Rough set approaches to rule induction from
incomplete data, in: Proceedings of the IPMU’2004, vol. 2, 2004, pp. 923–930.
[28] X.-H. Gu, D.-B. Hou, Z.-K. Zhou, D. Yu, Rough set based on modified chimerge
algorithm and its application, in: Proceedings of Machine Learning and Cyber-
netics, 2005, pp. 1004–1008.
[29] J.W. Guan, D.A. Bell, Rough computational methods for information, Artificial
Intelligence 105 (1998) 77–103.
[30] D. Haiying, Z. Yubo, X. Junyi, Hiearchical fault diagnosis for substation based on
rough set, Proceedings of the Powercon 4 (2002) 2318–2321.
[31] A.-E. Hassanien, Rough set approach for attribute reduction and rule generation:
a case of patients with suspected breast cancer, Journal of the American Society
for Information Science and Technology 55 (11) (2004) 954–962.
[32] A.-E. Hassanien, J.M.H. Ali, Enhanced rough sets rule reduction algorithm for
classification digital mammography, intelligent system journal, Freund & Pett-
man 13 (2) (2004).
[33] T.-P. Hong, L.-H. Tseng, S.-L. Wang, Learning rules from incomplete training
examples by rough sets, Expert Systems with Applications 22 (2002) 285–293.
[34] T.-P. Hong, T.-T. Wang, S.-L. Wang, Mining Fuzzy b-Certain and b-Possible Rules
from Quantitative Data based on the Variable Precision Rough Set Model, Expert
Systems with Applications, Elsevier Ltd., 2005.
[35] J. Hopfield, Neural networks and physical systems with emergent collective
computational abilities, Proceedings of the National Academy of Sciences U S A
(1982) 2554–2558.
[36] K. Hu, Y. Lu, C. Shi, Sampling for Approximate Reduct in Very Large Datasets,
http://www.lakecloud.xiloo.com/pkaw2000, http://www.citeseer.ist.psu.edu/
587308.html, 2000.
[37] X. Hu, Using rough sets theory and database operations to construct a good
ensemble of classifiers for data mining applications, Proceedings of ICDM (2001)
233–240.
[38] X. Hu, N. Cercone, Discovering maximal generalized decision rules through
horizontal and vertical reduction, Computational Intelligence 17 (4) (2001).
[39] X. Hu, T.Y. Lin, J. Jianchao, A new rough sets model based on database systems,
Fundamenta Informaticae (2004) 1–18.
[40] P. Jaganathan, K. Thangavel, A. Pethalakshmi, M. Karnan, Classification rule
discovery with ant colony optimization and Quickreduct Algorithm, in: Proceed-
ings of Intelligent Optimization Modeling, Allied Publishers, 2006.
[41] P. Jaganathan, K. Thangavel, A. Pethalakshmi, M. Karnan, Classification rule
discovery with ant colony optimization and improved Quick Reduct algorithm,
Lecture Notes in Engineers and Computer Scientists, Hong Kong, 2006, pp. 286–
291.
[42] I. Jagielska, C. Matthews, T. Whitfort, An investigation into the application of
neural networks, fuzzy logic, genetic algorithms and rough sets to automated
knowledge acquisition for classification problems, Neurocomputing 24 (1999)
37–54.
[43] R. Jensen, Q. Shen, Fuzzy-rough sets for descriptive dimensionality reduction, in:
Proceedings of the 11th International Conference on Fuzzy Systems, 2002, pp.
29–34.
[44] R. Jensen, Q. Shen, Finding rough set reducts with ant colony optimization, in:
Proceedings of the UK Workshop on Computational Intelligence, 2003.
[45] R. Jensen, Q. Shen, Fuzzy-rough attribute reduction with application to web
categorization, Fuzzy Sets and Systems 141 (3) (2004) 469–485.
[46] R. Jensen, Q. Shen, Semantics-preserving dimensionality reduction: rough and
fuzzy-rough-based approaches, IEEE Transactions on Knowledge and Data Engi-
neering 16 (12) (2004).
[47] R. Jensen, Q. Shen, Fuzzy-rough data reduction with ant colony optimization,
Fuzzy Sets and Systems 149 (2005) 5–20.
[48] L. Jiye, X. Zongben, Uncertainty measures of roughness of knowledge and rough
sets in incomplete information systems, in: Proceedings of the 3rd World
Congress on Intelligent Control and Automation, 2000, pp. 2526–2529.
[49] G.H. John, R. Kohavi, K. Pfleger, Irrelevant Features and the Subset Selection
Problem, in: Proceedings of 11th International Conference on Machine Learning,
1994, pp. 121–129.
[50] M. Karnan, K. Thangavel, A. Pethalakshmi, P. Jaganathan, Performance analysis of
classifiers, in: Proceedings of Intelligent Optimization Modeling, Allied Publish-
ers, 2006.
[51] L.P. Khoo, S.B. Tor, L.Y. Zhai, A rough set-based approach for classification and
rule induction, International Journal of Advanced Manufacturing Technology 15
(1999) 438–444.
[52] K. Kira, L.A. Rendell, The feature selection problem: traditional methods and a
new algorithm, in: Proceedings of AAAI, MIT Press, 1992, pp. 129–134.
[53] J. Komorowski, A. Ohrn, Modelling prognostic power of cardiac tests using rough
sets, Artificial Intelligence in Medicine 15 (1999) 167–191.
[54] S. Krishnaswamy, S.W. Loke, A. Zaslavsky, Estimating computation times of data-
intensive applications, IEEE Distributed Systems Online 5 (4) (2004).
11
[55] M. Kryszkiewicz, K. Cichon, Towards scalable algorithms for discovering rough
set reducts, Transactions on Rough Sets 1, LNCS 3100 (2004) 120–143.
[56] A. Kusiak, Decomposition in data mining: an industrial case study, IEEE Transac-
tions on Electronics Packaging Manufacturing 23 (4) (2000) 345–353.
[57] A. Kusiak, J.A. Kern, K.H. Kernstine, B.T.L. Tseng, Autonomous decision-making: a
data mining approach, IEEE Transactions on Information Technology in Biome-
dicine 4 (4) (2000) 274–284.
[58] A. Kusiak, K.H. Kernstine, J.A. Kern, K.A. McLaughlin, T.L. Tseng, Data mining:
medical and engineering case studies, Proceedings of the Industrial Engineering
Research (2000) 1–7.
[59] R. Latkowski, M. Mikolajczyk, Data decomposition and decision rule joining for
classification of data with missing values, Transactions on Rough Sets 1, LNCS
3100 (2004) 299–320.
[60] Y. Leung, W.Z. Wu, W.-X. Zhang, Knowledge acquisition in incomplete informa-
tion systems: a rough set approach, European Journal of Operational Research
168 (2006) 164–180.
[61] J. Li, N. Cercrone, Empirical Analysis on the Geriatric Care Data Set Using Rough
Sets Theory, Tech. Report, CS-2005-05, 2005.
[62] K. Li, Y.-S. Liu, Rough set based attribute reduction approach in data mining, in:
Proceedings of the First International Conference on Machine Learning and
Cybernetics, 2002.
[63] R. Li, Z.-O. Wang, Mining classification rules using rough sets and neural net-
works, European Journal of Operational Research 157 (2004) 439–448.
[64] Y. Li, S.C.-K. Shiu, S.K. Pal, J.N.-.K. Liu, A rough-set based CBR approach for feature
and document reduction in text categorization, in: Proceedings of the 3rd
International Conference on Machine Learning and Cybernetics, 2004.
[65] Y.-R. Li, J.-P. Jiang, The integrated of rough sets theory, fuzzy logic and genetic
algorithms for multisensor fusion, in: Proceedings of the American Control
Conference, 2001.
[66] T.Y. Lin, Granular computing: from rough sets and neighbourhood systems to
information granulation and computing with words, in: Proceedings of Eur-
opean Congress on Intelligent Techniques and Soft Computing, 1997.
[67] T.Y. Lin, N. Cercone, Rough sets and Data Mining: Analysis of Imprecise Data,
Kluwer Academic Publishers, 1997.
[68] H. Liu, H. Motoda, Feature Extraction Construction and Selection: A Datamining
Perspective, Kluwer Academic Publishers, 1998.
[69] P.K. Maji, A.R. Roy, An application of soft sets in a decision making problem, An
International Journal of Computers and Mathematics with Applications 44
(2002) 1077–1083.
[70] B. Mak, T. Munakata, Rule extraction from expert heuristics: a comparative study
of rough sets with neural networks and ID3, European Journal of Operational
Research 136 (2002) 212–229.
[71] H. Midelfart, J. Komorowski, K. Norsett, F. Yedetie, A.K. Sandwick, A. Laegreid,
Learning rough set classifiers from gene expressions and clinical data, Funda-
menta Informaticae 53 (2002) 155–183.
[72] S. Mitra, P. Mitra, S.K. Pal, Evolutionary modular design of rough knowledge-
based network using fuzzy attributes, Neurocomputing 36 (2001) 45–66.
[73] M. Modrzejewski, Feature selection using rough sets theory, in: Proceedings of
the 11th International Conference on Machine Learning, 1993, pp. 213–226.
[74] A. Muir, I. Duntsch, G. Gediga, Rough set data representation using binary
decision diagrams, RACSAM, Rev. R. Acad. Cien. Serie A. Mat 98 (1) (2004)
197–211.
[75] H. Nakayama, Y. Hattori, R. Ishii, Rule extraction based on rough set theory and
its application to medical data analysis, Proceedings of IEEESMC’99 5 (1999)
924–929.
[76] Z. Pawlak, Rough Sets, International Journal of Computer and Information
Sciences 11 (5) (1982) 341–356.
[77] Z. Pawlak, Rough Sets: Theoretical Aspects and Reasoning about Data, Kluwer
Academic Publishers, Dordrecht, 1991.
[78] W. Pedrycz, F. Gomide, An Introduction to Fuzzy Sets: Analysis and Design, The
MIT Press, 1998.
[79] A. Pethalakshmi, K. Thangavel, Rule identification in HIV data using rough set
strategies, in: Proceedings of Computing and Mathematical Modeling, Narosa
Publishing House, 2005, pp. 25–34.
[80] A. Pethalakshmi, K. Thangavel, P. Jaganathan, M. Karnan, Attribute reduction for
decision system with multiple decision attributes using rough set theory, in:
Proceedings of Intelligent Optimization Modeling, Allied Publishers, 2006.
[81] N.H. Phuong, L.L. Phuong, P. Santiprabhob, B.D. Baets, Approach to generating
rules for expert systems using rough set theory, Proceedings of IEEE (2001) 877–
882.
[82] D. Pyle, Data Preparation for Data Mining, Morgan Kauffman Publishers, 1999.
[83] F. Questier, I.A. Rollier, B. Walczak, D.L. Massart, Application of rough set theory
to feature selection for unsupervised clustering, Chemometrics and Intelligent
Laboratory Systems 63 (2002) 155–167.
[84] RSES: Rough Set Exploration System, http://logic.mimuw.edu.pl/rses/.
[85] G. Ruhe, Rough set-based data analysis in goal-oriented software measurement,
Proceedings of METRICS (1996) 10–19.
[86] A.B.M. Salem, M. Roushdy, S.A. Mahmoud, Mining patient data based on rough
set theory to determine thrombosis disease, International Journal on Artificial
Intelligence and Machine Learning 05 (1) (2004) 23–27.
[87] Q. Shen, A. Chouchoulas, A modular approach to generating fuzzy rules with
reduced attributes for the monitoring of complex systems, Engineering Applica-
tions of Artificial Intelligence 13 (3) (2002) 263–278.
[88] Q. Shen, A. Chouchoulas, A rough-fuzzy approach for generating classification
rules, Pattern Recognition 35 (2002) 2425–2438.
12 K. Thangavel, A. Pethalakshmi / Applied Soft Computing 9 (2009) 1–12
[89] Q. Shen, R. Jensen, Selecting informative features with fuzzy-rough sets and its
application for complex systems monitoring, Pattern Recognition 37 (2004)
1351–1363.
[90] R. Slowinski, J. Stefanowski, Rough classification in incomplete information
systems, Mathematical Computer Modelling 12 (10/11) (1989) 1347–1357.
[91] J. Starzyk, D.E. Nelson, K. Sturtz, Reduct generation in information systems,
Bulletin of International Rough Set Society 3 (1998) 19–22.
[92] R.W. Swiniarski, Rough sets methods in feature reduction and classification,
International Journal on Applied Mathematics and Computer Science 11 (3)
(2001) 565–582.
[93] R.W. Swiniarski, A. Skowron, Rough set methods in feature selection and
recognition, Pattern Recognition Letters 24 (2003) 833–849.
[94] R.R. Tan, Rule-based life cycle impact assessment using modified rough set
induction methodology, Environmental Modelling and Software 20 (2005) 509–
513.
[95] F.E.H. Tay, L. Shen, Fault diagnosis based on rough set theory, Engineering
Applications of Artificial Intelligence 16 (2003) 39–43.
[96] K. Thangavel, Q. Shen, A. Pethalakshmi, Application of clustering for feature
selection based on rough set theory approach 6 (1) (2005) 19–27.
[97] K. Thangavel, A. Pethalakshmi, Feature selection for medical database using
rough system, International Journal on Artificial Intelligence and Machine
Learning 6 (1) (2005) 11–17.
[98] K. Thangavel, M. Karnan, A. Pethalakshmi, Performance analysis of rough reduct
algorithms in mammogram, International Journal on Global Vision and Image
Processing 5 (8) (2005) 13–21.
[99] K. Thangavel, P. Jaganathan, A. Pethalakshmi, M. Karnan, Effective classification
with improved quick reduct for medical database using rough system, Bioinfor-
matics and Medical Engineering 05 (1) (2005) 7–14.
[100] K. Thangavel, A. Pethalakshmi, P. Jaganathan, A comparative analysis of feature
selection algorithms based on rough set theory, International Journal of Soft
Computing 1 (4) (2006) 288–294.
[101] K. Thangavel, A. Pethalakshmi, P. Jaganathan, A novel reduct algorithm for
dimensionality reduction with missing values based on rough set theory,
International Journal on Soft Computing 1 (2) (2006) 111–117.
[102] K. Thangavel, M. Karnan, P. Jaganathan, R. Sivakumar, A. Pethalakshmi, Com-
puter-Aided Diagnosis: Automatic detection of Microcalcifications in Mammo-
graphic Images using Soft Computing, Lecture Notes in Engineers and Computer
Scientists, Hong Kong, 2006, pp. 280–285.
[103] K. Thangavel, M. Karnan, P. Jeganathan, A. Pethalakshmi, R. Sivakumar, G.
Geetharamani, Ant colony algorithms in diverse combinational optimization
problems—a survey, International Journal on Automatic Control and System
Engineering 6 (1) (2006) 7–26.
[104] K. Thangavel, A. Pethalakshmi, Performance analysis of accelerated Quickreduct
algorithm, Proceedings of International Conference on Computational Intelli-
gence and Multimedia Applications; IEEE Xplore 2.0 2 (2007) 318–322.
[105] The ROSETTA homepage (http://www.idi.ntnu.no/aleks/rosetta/), Norwegian
University of Science and Technology, Department of Computer and Information
Science.
[106] Y.-C. Tsai, C.-H. Cheng, J.-R. Chang, Entropy-based fuzzy rough classification
approach for extracting classification rules, Expert Systems with Applications
(2005) 1–8.
[107] T.-L. (Bill)Tseng, Y. Kwon, Y.M. Ertekin, Feature-based rule induction in machin-
ing operation using rough set theory for quality assurance, in: Robotics and
Computer-Integrated Manufacturing, Elsevier Publishers, 2005.
[108] S. Tsumaoto, Mining diagnosis rules from clinical databases using rough sets and
medical diagnostic model, Information Sciences 162 (2004) 65–80.
[109] L. Vesa, L. Mikko, K. Hannu, Network traffic classification using rough sets, in:
Third WSEAS Symposium on Mathematical Methods and Computational Tech-
niques in Electrical Engineering, 2001.
[110] L. Vesa, L. Timo, L. Mikko, K. Hannu, A comparison of fuzzy C-means clustering
and rough sets based classification in network data analysis, in: Proceedings of
WSEAS NNA-FSFS-EC, Interlaken, Switzerland, 2002.
[111] S. Vinterbo, A. Ohrn, Minimal approximate hitting sets and rule templates,
International Journal of Approximate Reasoning 25 (2000) 123–143.
[112] F.H. Wang, On acquiring classification knowledge from noisy data based on
rough set, Expert Systems with Applications 29 (2005) 49–64.
[113] Y. Wang, M. Ding, C. Zhou, T. Zhang, A hybrid method for relevance feedback in
image retrieval using rough sets and neural networks, International Journal of
Computational Cognition 3 (1) (2005).
[114] R.H. Warren, J.A. Johnson, G.H. Huang, Application of rough sets to environmental
engineering models, Transactions on Rough Sets 1, LNCS 3100 (2004) 356–374.
[115] J.-M. Wei, Rough set based approach to selection of node, International Journal of
Computational Cognition 1 (2) (2003) 25–40.
[116] S.Z. Wilk, R. Slowinski, W. Michalowski, S. Greco, Supporting triage of children
with abdominal pain in the emergency room, European Journal of Operational
Research 160 (3) (2005) 696–709.
[117] L. Yidong, Z. Ling, L. Lianchen, Reducing inconsistent rules based on irregular
decision table, Tsinghua Science and Technology 9 (1) (2004) 45–50.
[118] X. Yin, Z. Zhou, N. Li, S. Chen, An Approach for Data Filtering Based on Rough Set
Theory LNCS 2118, Springer-Verlag, Berlin, 2001, pp. 367–374.
[119] J. Zaluski, R. Szoszkiewicz, J. Krysinski, J. Stefanowski, Rough set theory and
decision rules in data analysis of breast cancer patients, Transactions on Rough
Sets 1, LNCS 3100 (2004) 375–391.
[120] X. Zeng, Y. Zhan, Development of a noise sources classification system based on
new method for feature selection, Applied Acoustics 66 (2005) 1196–1205.
[121] J. Zhang, J. Wang, D. Li, H. He, J. Sun, A New Heuristic Reduct Algorithm Base on
Rough Sets Theory, LNCS 2762, Springer-Verlag, 2003, pp. 247–253.
[122] M. Zhang, J.T. Yao, A rough sets based approach to feature selection, fuzzy
information, Processing NAFIPS’04 1 (2004) 434–439.
[123] N. Zhong, A. Skowron, A rough set-based knowledge discovery process, Inter-
national Journal of Applied Mathematics and Computer Science 11 (3) (2001)
603–619.
[124] N. Zhong, J. Dong, S. Ohsuga, Using rough sets with heuristics for feature
selection, Journal of Intelligent Information Systems 16 (2001) 199–214.
[125] W. Zhu, F.-Y. Wang, Reduction and axiomization of covering generalized rough
sets, Information Sciences 152 (2003) 217–230.