WIXLIB

82Journal of Industrial Engineering, University of Tehran, Special Issue, 2011Phase 1: Business and data understandingBusiness statusBusiness problemData mining goalsData collectionand selectionPhase 3: ModelingFind optimum Kwith Dunn indexEstimating CLVscore of clustersClustering customerbased on RFM parameterSegmenting customers to 4groups based on similarity inCLV score and dedicatingrank to themPhase 2: Data preparing and preprocessingExtracting data: Transaction Data(Recency (R),Frequency(F), andMonetary (M) parameters Demographic Data(Age, Sex, and BranchZone) Behavioral Data(account type, cardownership status)ScalingR, F, MparametersMergingTransaction,Demographicand ProductownershipdataPhase 4: EvaluationEvaluating the results with metricsBalancing imbalancedsegmentsExtract rules basedon demographic andbehavioralpredictors fordifferent segmentsExtract rules based ontransaction predictors fordifferent segment with AccuracyPrecisionRecallF-measureclassification methodEvaluating that which business objectives aremet by the proposed work (in marketingstrategy scope)Figure 2: Structure of Research methodologyData mining methods are two categories:descriptive and predictive. Descriptivemethod such as classification and predictivemethod such as clustering are used in currentissue. Ngai et al. compare classification andclustering models as follow [25]:Classification aims at building a model topredict future customer behaviors throughclassifying database records into a number ofpredefined classes based on certain criteria,whereas clustering is the task of segmentinga heterogeneous population into a number ofmore homogenous clusters. It is different toclassification in that clusters are unknown atthe time the algorithm starts. In other words,there are no predefined clusters. Commontools used for classification are neuralnetworks, decision trees and if then-elserules [12] and common tools for clusteringinclude neural networks and discriminationanalysis [26].In CLV context, both of these modelsused in academic research, Baesens et al.used classification Bayesian networkclassifier in [27] and Drew et al. employedclustering neural network [28]. In this study,K-means algorithm as a clustering methodand several decision tree algorithms asclassification methods are used.

Analyzing Customer Segmentation .2. Methodology2.1. Crisp methodologyThere are different methodologies forimplementing data mining projects but oneof the powerful methods is CRISP (CRossIndustry Standard Process for Data Mining)methodology [29]. As a process model,CRISP provides an overview of the datamining life cycle. CRISP uses six phases todescribe the process from gathering businessrequirements to deploying the results [29]: Business Understanding: This phasetypically involves gathering requirementsand meeting with expert personnel todetermine goals rather than working withdata. Dataunderstanding:Thedataunderstanding phase of CRISP involvestaking a closer look at the data availablefor mining. This phase includescollecting initial data, describing data,exploring data, and verifying dataquality. Data preparation: Data preparation is oneof the most important and often timeconsuming aspects of data miningprojects and includes selecting data,cleaning data, constructing new data, andintegrating data. Modeling: The data which was spenttime preparing are ready to bring intodata mining algorithms, and the resultsbegin to shed some light on the businessproblem posed. Selecting modelingtechniques, generating a test design,building the models, and assessing themodel construct this phase. Evaluation: In this phase, evaluating theresults, review process, and determiningthe next steps are done. Deployment: Deployment is the processof using the new insights to makeimprovements within the organization.2.2. Research methodologyThe CRISP methodology is flexible andcan be customized easily. With regards tothis methodology, a research methodologywas proposed to ranking customers based onsegmenting customer with customer value83driver parameters. Figure 2 illustrateresearch methodology of this study.Phase 1: Business and data understanding.The case study concerns a private bank inIran. It was established by a consortium ofindustrial, construction and investmentcompanies, with the aim of providingflexible financial services to the burgeoningIranian private sector. It has since grown tooffer a full range of retail and commercialbanking solutions, serving individualcustomers, small to medium sized businessesand large corporations and publicinstitutions, both online and through ourextensive national network of branches. Datamining department is a new department inthis financial institute. This department triesto construct data warehouse from its hugedata bases to replace analyzing on data basedon OLTP (OnLine Transaction Process)system with OLAP (OnLine AnalyticalProcess).The most important goal of thisorganization is presentation of best servicesto customers. But the problem is that there isno taxonomy of customers based on theirvalue in the bank. The main objective ofcurrent study is analysis on the customers’data to discover similar pattern betweendifferent segments of customer based ontheir equity and then assigning each segmentof customer a rank based on their value, thenextracting rule set for different segment leadto identifying high-value customers. Theresults of the study can then be used by thebusiness to establish marketing and CRMstrategy for different group of customers.For this purpose, two years data oftransactions, approximately five millionsrecords for 50,000 customers, have beencollected with simple stochastic samplingmethod. The data can be categorized , transaction information, andproduct ownership information. Table 1explains some information about data.Phase 2: Data preparing and preprocessing.

84Journal of Industrial Engineering, University of Tehran, Special Issue, 2011Extracting data, scaling R, F, M parameters,merging transaction, demographic, andproduct ownership data are sub phase of datapreparing and preprocessing phase.There are transaction data in detail whichinclude type of transaction, date oftransaction, monetary value of transaction(positive or negative), and type of deposit.For extracting R, F, and M parameters ofeach customer we aggregate records basedon customer ID. In each aggregated record,there is a unique customer with her/hisnumber of monetary transaction as Fparameter, latest transaction date as Rparameter and total amount of money in allof her/his deposit account at the end of thecertain period (2 years) as M parameter.Then we scale each of these three parametersin five scales, very low, low, medium, high,and very high scale. For the scaling Rparameter, first sort the data based on Rattribute by ascendant order then partitionthe customers transaction dataset into 5partitions. Customers in the first partitionhave lowest value of Recency and theirRecency value named very low, secondpartition named low, third partition hasmedium scale, fourth is high scale andfinally fifth partition named very high. Forthe scaling F and M parameters the sameprocedure must be done.In the socio-demographic data, there is afield that called Branch zone, this attribute isderived from branch code. Tehran is dividedinto 22 zones (1 to 22) and has 95 branches,in the other hand, 111 branches aredistributed to the other cities. These cities aredivided into 6 zones (23 to 28) based ondefined supervision zone in the bank.In last step of this phase, we integrate datawith merging socio-demographic, productownership data and the result table of RFMaggregation. At the earliest, we have 50000customers’ data, but in the end of this phasewe have a new aggregated data that includeinformation (socio-demographic, productownership, and aggregated transactioninformation) about 34000 customers.Table 1: Data category and descriptionData demographicinformationBranch vingaccountShort termtimedepositaccountsLong termtimedepositaccountsDebit cardownershipIs a range of 2 to 880 for men and 1 forwomenThere are 28 zonesthat 22 of them are inTehran. It is derivedfrom Branch code ofeach customer.Latesttransactiondate during 2 yearsNumber of monetarytransactionmadewithin 2 yearsTotal amount ofmoney in all of thecustomer’s depositsin the end of theperiodTotal numbers ofproduct1 for ownership 0for non ownership1 for ownership 0for non ownership1 for ownership 0for non ownership1 for ownership 0for non ownership1forcardownership 0 fornon card ownership.Scaling5 scoreScaling nameVery high (VH)Table 2: Scaling RFM parametersRecencyFrequency2010/1/6- 2010/1/10[126,16968)4 score3 score2 score1 scoreHigh (H)Medium (M)Low (L)Very low (VL)2009/12/23- 2010/1/62009/11/24- 2009/12/232009/4/1- 2009/11/242008/1/9- 79 35,424,208,1462,027,705 - 26,913,379253,333- 2,027,70518.213 - 253,3330 - 18,213

Analyzing Customer Segmentation .Phase 3 and 4: Modeling and evaluationAs it is mentioned in first phase, the goalof using data mining tools in this study isranking customers based on ion is done in two ways: customerneed-based and customer value-basedsegmentation [7, 8, 12, and 30]. Inparticular,customerneeds-basedsegmentation must ultimately drive CRMprograms even though a customerprofitability segmentation analysis can serveas a good starting point for efficient programdevelopment [31].For clustering customers, there aredifferent algorithms which can be used. Avariety of factors need to be consideredwhen deciding which type of clusteringtechnique to use [32]. One of the mostcommon iterative algorithms is the K-meansalgorithm [33] which can be used for a widevariety of data types. It is also quiteefficient, even though multiple runs areoften performed. K-means has troubleclustering data that contains outliers. Outlierdetection and removal can help significantlyin such situations. Finally, K-means isrestricted to data for which there is a notionof a center (centroid) [32]. At first step ofmodeling phase, we cluster customers basedon customer value drivers (recency,frequency and monetary), with K-meansalgorithm. There are some indices forfinding the optimum K (the number ofclusters) that reviewed in [34]. We usedDunn index in this study [35]. The maingoal of this measure is to maximizeintercluster distances (distance betweendifferent clusters), whilst minimizingintracluster distances (distance betweenmembers of a cluster) [34]. For any,wherepartitionthrepresents the i cluster of such partition,the Dunn indices, D, is defined as inequation (1):85(1)Where σ(ci, cj) defines the distancebetween clusters ci , cj(intercluster (ci) represents thedistance), andintracluster distance of cluster ci or the sizeof the cluster cj , and k is the total number ofclusters. In this study:(2)(3)where and are the centroid of clusterand. Thus large Value of Dcorresponds to good clusters. Therefore, thenumber of cluster that maximizes D is takenas the optimal number of clusters, c. In thisstudy optimum number of clusters (K) is 9which calculated by Dunn index (Figure 3).Figure 3: Optimum K with Dunn Second, based on centroid point of eachcluster and information of table 2, R, F, andM score are dedicated to each cluster, as it isshown in table 3.

86Journal of Industrial Engineering, University of Tehran, Special Issue, 2011Table 3: Clustering customers with RFM parametersc1c2c3No. H - VH - 4641,384VH - VH - VHVL - L - Mc7209242009/12/2912098,573,791H - VH - VHc826632009/10/10576,525,315L - H - Hc925312009/03/06231,887,606VL - L - MclusterRFMRFM 0,457934,639,633649,263L - M - MH - VH - VHVL - VL - MTable 4: RFM Score of clusterscluster% of customerR ScoreF scoreM ScoreRFM 6ParameterTable 5: Weights of RFM parametersRFMWeight0.0870.3450.653Table 6: Calculating CLV with approach 2ClusterNormal RNormal FNormal 340.0900.0770.0514114234

Analyzing Customer Segmentation .87RFM score of each cluster can becalculated with equation (4), as follow:Whererefers to recency score ofis frequency score ofcluster ci,cluster ci, andi is monetary scoreof cluster ci (table 4).Cluster 2 and 7 with 20929 memberswhich include 60 percents of all customershave 14 RFM score, other hand, C4 and C5altogether have 395 members which includejust one percent of all customers have 15RFM score, as it is shown in figure 4.original data [32]. Suppose that minA andmaxA are the minimum and maximumvalues of an attribute, A. Then min-maxnormalization maps a value, v, of A to in] bythe range of [computing:(6)In current study the new range is [0, 1].According to the assessments obtained bythe AHP method with regard to expert’sinsight of the bank, the relative weights ofthe RFM variables are mentioned in table 5.In table 6 the normal R, F and M valuesfor each cluster's centroid and CLV of themis shown.Figure 4: Diagram of customers' RFM scoreAfter that, we calculate CLV score of eachcluster based on weighted RFM analysis asfollow in equation (5):(5)whererefers to normal recency ofcluster ,is the weight of recency,is normalized frequency,is theweight of frequency,is normalizedmonetary, andis the weight ofmonetary. In this study, min-maxnormalization method is used fornormalizing data. Min-max normalizationperforms a linear transformation on theFigure5: Distribution of customers in segmentsIn general we have four segments ofcustomers (as it is shown in last column oftable 6). For analyzing customers’transaction behavior based on their valuedriver parameters, classifier algorithm canbe used to extract rules. In fact, we extractrules for explanation of clear knowledgeabout different segments of customers withdifferent ranks. In current study, severalclassifiers are used and PART (ProjectiveAdaptive Resonance Theory) algorithm ischosen for extracting rules [36] because ofits best metrics results (see table 8). PARTuses separate-and-conquer and builds apartial C4.5 decision tree in each iteration

88Journal of Industrial Engineering, University of Tehran, Special Issue, 2011and makes the best leaf into a rule. Thecomparison among decision tree classifiersand rule set classifiers is shown in table 8and the rule set of PART decision ruletechnique is presented in list 1.Metrics are used to guide the data miningalgorithms and to evaluate the results of datamining. Recall and precision are two widelyused metrics employed in applicationswhere successful detection of one of theclasses is considered more significant thandetection of the other classes [32]. Precisionand recall can be summarized into anothermetric known as the F-measure:Where is: TP corresponds to the number ofpositive examples correctly predicted bythe classification model FN corresponds to the number ofpositive examples wrongly predicted asnegative by the classification model FP corresponds to the number ofnegative examples wrongly predicted aspositive number of negative exampleswrongly predicted as positive by theclassification model TN corresponds to the number ofnegative examples correctly predicted bythe classification modelTable 8: Comparison among 9.9498.8210.9990.97710.9990.98810.9990.983To discover the pattern betweencustomers’ value base segmentation (the lastcolumn of table 6) as target field and theirsocio-demographic and product ownershipattributes as predictors, rule set algorithmcan be used to extract rules. The algorithmwhich is exploited in current study is PARTdecision rule. Because of unbalancingproblem among the four segments, all of therules terminate in segment 2 with 60percents of all customers. In a conceptlearning problem, the data set is said topresent a class imbalance if it contains manymore examples of one class than the other.Such a situation poses challenges for typicalclassifiers such as decision tree inductionsystems or multi-layer perceptrons that aredesigned to optimize overall accuracywithout taking in to account the relativedistribution of each class. As a result, theseclassifiers tend to ignore small classes whileconcentrating on classifying the large onesaccurately. There are some of the methodsdeveloped for handling the class imbalanceproblem [37]. Sampling-based methods andalternative metrics besides accuracy (i.e.recall and precision) introduced in this side.Data-based techniques for sampling includeundersampling, oversampling, and a hybridof both approaches. In the current study, weused hybrid one to reduce overfittingproblem which is the result of oversamplingmethod and omitting certain importantrecords that is the result of undersamplingmethod.

Analyzing Customer Segmentation .As it is mentioned prior, predictors fieldsinclude socio-demographic and productownership attributes and the target field isthe segment number. So, we run PARTdecision rule two times, first with productownership data as predictors and secondwith socio-demographic data. Survey onresults shows that socio-demographic dataof customers are not good predictors of thesegment of customers, because the accuracyof the rules is about 37 percents, that is verylow. But product ownership data could beproper predictors for customer basesegmentation. Before sampling, all of therules terminated in segment 2 with 60percents of all customers but after balancingdata, we could extract a rule set for productownership variables and the segmentation ofthe customers, which cover all segments(see list 2). The discovered knowledge issummarized in table 9. The content of table9 show that Current account, the totalnumber of product ownership and long termtime deposit accounts have a positiverelation with CLV score of customers. Alsosocio-demographic data such as age, sex andbranch zone have no relation with customervalue score.2.2.1. Segmentations ranking analysis andproposed strategySegment 1: First segment includes themost valuable customers of the financialinstitute. This group is constructed from C2,C4, and C5 which have high CLV value.These customers are labeled “goldcustomers” and just one percent ofcustomers belong to this segment. The bankmust have special program and strategy forthem. Private banking that is one of the threescopes of banking marketing strategies, mustbe consider for this segment. The bankshould give consultant services to this groupfor managing their property.89Segment 2: Cluster 7 with 60 percents ofcustomers’ contribution constructs thissegment. This group contains customerswho averagely refer to the bank one time ina week and the last transaction date of themis about ten days ago and average totalmonetary value of them is very high score ofmonetary scoring in retail banking scope. Sothis segment includes valuable customers ofthe bank and the insti

of customer with different CLV rank. The rest of this study is organized as follows. Section 2 outlines the background and reviews related work on customer lifetime value, CLV divisions and classifications, RFM analysis, and data mining definitions and techniques in customer segmentation application. Section 3 describes the research methodology and