(IJCSIS) International Journal of Computer Science and Information Security,Vol. 9, No. 4, April 2011
Dynamic Rough Sets Features Reduction
Walid MOUDANI
1
, Ahmad SHAHIN
2
, Fadi CHAKIK
2
, and Félix Mora-Camino
3
1
Lebanese University, Faculty of Business, Dept. of Business Information System, Lebanon
2
LaMA – Liban, Lebanese University, Lebanon
3
Air Transportation Department, ENAC, 31055 Toulouse, France
Abstract
—Nowadays, and with the current progress intechnologies and business sales, databases with large amount of data exist especially in retail companies. The main objective of this study is to reduce the complexity of the classificationproblems while maintaining the prediction classification quality.We propose to apply the promising technique Rough Set theorywhich is a new mathematical approach to data analysis based onclassification of objects of interest into similarity classes, whichare indiscernible with respect to some features. Since somefeatures are of high interest, this leads to the fundamentalconcept of “Attribute Reduction”. The goal of Rough set is toenumerate good attribute subsets that have high dependence,discriminating index and significance. The naïve way of is togenerate all possible subsets of attribute but in high dimensioncases, this approach is very inefficient while it will require
12
d
iterations. Therefore, we apply Dynamic programmingtechnique in order to enumerate dynamically the optimal subsetsof the reduced attributes of high interest by reducing the degreeof complexity. Implementation has been developed, applied, andtested over a 3 years historical business data in Retail Business.Simulations and visual analysis are shown and discussed in orderto validate the accuracy of the proposed tool
Keywords- Data Mining; Business Retail; Rough Sets; Attribute Reduction; Classification; Dynamic Programming.
I.
I
NTRODUCTION
Retail Business (RB) Company looks for increasing its benefitby providing all facilities services to its customers. Theestimated benefits amount to several millions of dollars whenthe Retail Business Company organizes and offers to itscustomers the most related items. The RB Company stores andgenerates tremendous amounts of raw and heterogeneous datathat provides rich fields for Data Mining (DM) [1, 2]. Thisdata includes transactions Details (customers/providers)describing the content such as items, quantity, date, unit price,reduction, and other events such as the holidays, specialactivities, etc. Moreover, the profile of customers and theirfinancial transactions contribute in personalizing some specialservices to each customer. This leads the research communityto study deeply this field in order to propose a new solutionapproach for these companies. Moreover, these companiesshould analyze their business data in order to predict theappropriate services to be proposed to its customers. Thisapproach is one of the main objectives of the retailer company.In order to build such a non trivial model, many researcheswere carried out on the feasibility of using the DM techniques,which raised from the need of analyzing high volumes of datacollected by the retailer companies and related to differentkinds of transactions between the company and itscustomers/providers. Our contribution aims to reduce thecomplexity of the classification process by reducing thenumber of attributes that should be considered in order todiscover the fruitful knowledge required by decision makers of RB.The 1990s has brought a growing data glut problem tomany fields such as science, business and government. Ourcapabilities for collecting and storing data of all kinds have faroutpaced our abilities to analyze, summarize, and extractknowledge from this data [9]. Traditional data analysismethods are no longer efficient to handle voluminous datasets. How to understand and analyze large bodies of data is adifficult and unresolved problem. The way to extract theknowledge in a comprehensible form for the huge amount of data is the primary concern. DM refers to extractingknowledge from databases that can contain large amount of data describing decisions, performance and operations.However, analyzing the database of historical data containingcritical information concerning past business performance,helps to identify relationships which have a bearing on aspecific issue and then extrapolate from these relationships topredict future performance or behavior and discover hiddendata patterns. Often the sheer volume of data can make theextraction of this business information impossible by manualmethods. DM treats as synonym for another popularly usedterm, Knowledge Discovery in Databases. KDD is thenontrivial process of identifying valid, novel, potentiallyuseful and ultimately understandable patterns in data. DM is aset of techniques which allows extracting useful businessknowledge, based on a set of some commonly used techniquessuch as: Statistical Methods, Case-Based Reasoning, NeuralNetworks, Decision Trees, Rule Induction, Bayesian Belief Networks, Genetic Algorithms, Fuzzy Sets, Rough Sets, andLinear Regression [4, 36]. DM commonly used in a variety of domains such as: marketing, surveillance and fraud detectionin telecommunications, manufacturing process control, thestudy of risk factors in medical diagnosis, and customersupport operations through a better understanding of customers in order to improve sales.In commerce, RB is defined by buying goods or products inlarge quantities from manufacturers or importers, eitherdirectly or through a wholesaler, and then sells individualitems or small quantities to the general public or end usercustomers. RB is based on the sale of goods from fixed
1http://sites.google.com/site/ijcsis/ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 9, No. 4, April 2011
locations, these locations could be physical (shop or store)and/or virtual over the web. Retailing may include severaltypes of services that can go along with the sale, such asdelivery of goods, processing and tracking loyalty cardfunctionality. The process goes from buying products in largequantities from manufacturers, and then sells smallerquantities to the end-user. From a business perspective, DM ismainly used in the Customer Relationship Management(CRM) area, specifically marketing. DM today's applicationsprovide the tool for retailers or decision maker to get preciousknowledge that covers the requested field of interest and makesense of their customer data and apply it to business such as:the sales/marketing domain and other business-related areas[4]. It contributes to predict customer purchasing behavior andperform target marketing by using demographic data andhistorical information, to drive sales suggestions for alternateor related items during a purchase transaction, to identifyvaluable customers, allowing the CRM team to target them forretention, to point out potential long-term customers who canbe a potential target through marketing programs [36], toidentify people behavior who are likely to buy new productsbased on their item categories purchased, to assess theproducts which are bought together.This paper is organized as follows: in section 2, thebackground of DM and its relationship with RB is presentedand highlighted by specifying the main major problems facedby retailer. In section 3, we present the Rough Sets (RS)technique and the Rough Sets Attribute Reduction (RSAR)problem followed by a general overview of the literature and amathematical formulation. Therefore, in section 4, we present anew dynamic solution approach for the RSAR problem basedon the Dynamic Programming technique followed by a study of its complexity. In section 5, we describe our solution approachthrough a numerical example using some well-known datasetsfollowed by discussion and analysis of the results obtained.And finally, we ended by a conclusion concerning this new
approach and the related new ideas to be tackled in the future.
II.
R
OUGH
S
ET
T
HEORY
Pawlak has introduced the theory of RS which is anefficient technique for knowledge discovery in databases [33,34]. It is a relatively new rigorous mathematical technique todescribe quantitatively uncertainty, imprecision and vagueness.It leads to create approximate descriptions of objects for dataanalysis, optimization and recognition. It is shown to bemethodologically significant in the domains of ArtificialIntelligence and cognitive science, especially in respect of therepresentation and of the reasoning with imprecise knowledge,machine learning, and knowledge discovery. In RS theory, thedata is organized in a table called decision table. Rows of thedecision table correspond to objects, columns correspond toattributes, and class label indicates the class to which each rowbelongs. The class label is called as decision attribute, the restof the attributes are the condition attributes. Therefore, thepartitions/classes obtained from condition attributes are calledelementary sets, and those from the decision attribute(s) arecalled concepts. Let’s consider C for the condition attributes, Dfor the decision attributes, where
DC
, and
j
t
denotesthe
th
j
tuple of the data table. The goal of RS is to understandor construct rules for the concepts in terms of elementary sets,i.e., mapping partitions of condition attributes to partitions of decision attribute [41]. However, a RS is a formalapproximation of a crisp set in terms of a pair of sets whichgive the lower and the upper approximation of the original set.Once the lower and upper approximation is calculated, positive,negative, and boundary regions can be derived from theapproximation. Therefore, RS theory defines five regions basedon the equivalent classes induced by the attribute values. Lowerapproximation contains all the objects, which are classifiedsurely based on the data collected, Upper approximationcontains all the objects which can be classified probably,Negative region contains the set of objects that cannot beassigned to a given class, Positive region contains the objectsthat can be unambiguously assigned to a given class, while theBoundary is the difference between the upper approximationand the lower approximation which contains the objects thatcan be ambiguously (with confidence less than 100%) assignedto a given class.
A.
Elements of the rough sets
To illustrate clearly the RS technique, let’s consider the mainelements of RS theory. Let U be any finite universe of discourse. Let R be any equivalence relation defined on U,which partitions U. Here, (U, R) is the collection of allequivalence classes. Let
n
X X X
,,
21
be the elementary setsof the approximation space (U, R). This collection is known asknowledge base. Let A be a subset of U.Elementary sets:
m A
X X X R
,,
21
where
i
X
denote theelementary sets.(1)Concepts:
k Class
Y Y Y R
,,
21
where
i
Y
refer to concepts.(2)Lower approximation: Thus the lower approximation of aconcept is the set of those elementary sets that are containedwithin subset of the concept with probability of 1.
i j ji
A
Y X where X Y R
,)(
(3)Upper approximation: The upper approximation of a concept isthe set of those elementary sets that share some objects withthe concept (non-zero probability).
i j ji
A
Y X where X Y R
,)((4)Positive region: Thus the positive region of a concept is the setof those elementary sets that are subset of the concept. Positiveregion would generate the strongest rule with 100%confidence.
)()(
i Ai A
Y RY POS
(5)
2http://sites.google.com/site/ijcsis/ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 9, No. 4, April 2011
Boundary region: Thus the boundary region of a concept is theset of those elementary sets that have something to say aboutthe concept, excluding the positive region. It consists of thoseobjects that can neither be ruled in nor ruled out as members of the target set. These objects can be ambiguously (withconfidence less than 100%) assigned the class denoted by
i
Y
.Hence, it is trivial that if
A
BND
, then A is exact. Thisapproach provides a mathematical tool that can be used to findout all possible reducts.
)()()(
i Ai Ai A
Y RY RY BND
(6)Negative region: Thus the negative region of a concept is theset of those elementary sets that have nothing to say about theconcept. These objects cannot be assigned the class denotedby
i
Y
(their confidence of belonging to class
i
Y
is in fact 0%!)
)()(
i Ai A
Y RU Y NEG
(7)Concept Set: Concept set is the equivalence relation from theclass and elementary set are equivalence relation fromattributes. As mentioned above, the goal of the rough set is tounderstand the concept in term of elementary set. In order tomap between elementary set and concept, lower and upperapproximation must first defined. Then positive, boundary andnegative regions can be defined based on the approximationsto generate rules for categorization. Once the effect of subclassof concept is defined, the last step before rule generation is todefine the net effect on entire set of concepts. Given effect of subset of concept)(
i A
Y POS
, the net effect on entire set of concepts is defined as:
k ii A A
Y POSY POS
1
)()(
k ii A
Y BNDY BND
1
)()(
k ii A A
Y RY NEG
1
)()(
(8)Generating rules: There are two kinds of rules that can begenerated from the POS and the BND regions respectively. Forany
)(
j Ai
Y POS X
, we can generate a 100% confidence ruleof the form: If
i
X
then
j
Y
(or
ji
Y X
). For any)(
i Ai
Y BND X
we can generate a <100% confidence rule of the form: If
i
X
then
j
Y
(or
ji
Y X
), with confidence givenas:
i ji
X Y X conf
(9)Assessment a rule: As mentioned above, the goal of the RS isto generate a set of rules that are high in dependency,discriminating index, and significance. There are threemethods of assessing the importance of an attribute:-
Dependency:
How much does a class depends on A (subsetof attribute)
U classPOSclass
A A
)()(
(10)-
Discriminating Index:
Attributes A’s ability to distinguishbetween classes
U class BNDU class
A A
)()(
U class NEGclassPOS
A A
)()(
(11)-
Significance:
How much does the data depend on theremoval of A
)()()(
,,2,1,,2,1
classclassclass
Ad A A Ad A A A A
(12)Significance of A is computed with regard to the entire setof attributes. If the change in the dependency after removing Ais large, then A is more significant.
B.
Rough Set Based Attribute Reduction1)
Literature overview
Attribute or feature selection is to identify the significantfeatures, eliminate the irrelevant of dispensable features to thelearning task, and build a good learning model. It refers tochoose a subset of attributes from the set of original attributes.Attribute or feature selection of an information system is a keyproblem in RS theory and its applications. Usingcomputational intelligence tools to solve such problems hasrecently fascinated many researchers. Computationalintelligence tools are practical and robust for many real-worldproblems, and they are rapidly developed nowadays.Computational intelligence tools and applications have grownrapidly since its inception in the early nineties of the lastcentury [5, 8, 16, 24]. Computational intelligence tools, whichare alternatively called soft computing, were firstly limited tofuzzy logic, neural networks and evolutionary computing aswell as their hybrid methods [16, 40]. Nowadays, thedefinition of computational intelligence tools has beenextended to cover many of other machine learning tools. Oneof the main computational intelligence classes is GranularComputing [25, 40], which has recently been developed tocover all tools that mainly invoke computing with fuzzy andrough sets.However, some classes of computational intelligence tools,like memory-based heuristics, have been involved in solvinginformation systems and DM applications like other well-known computational intelligence tools of evolutionarycomputing and neural networks. One class of the promisingcomputational intelligence tools is memory-based heuristics,like Tabu Search (TS), which have shown their successfulperformance in solving many combinatorial search problems[10, 32]. However, the contributions of memory-basedheuristics to information systems and data mining applicationsare still limited compared with other computational
Identify applicable sponsor/s here.
(sponsors)
3http://sites.google.com/site/ijcsis/ISSN 1947-5500
Search History:
Result 00 of 00
00 results for result for
p.
Dynamic Rough Sets Features Reduction
Nowadays, and with the current progress in technologies and business sales, databases with large amount of data exist especially in retail companies. The main objective of this study is to reduce the complexity of the classification…
(More)
445
Reads
0
Readcasts
2
Embed Views
More From This User
Notes
Load more
