A Case Study of Applying SOM in Market Segmentation of Automobile Insurance Customers

Transcription

1 International Journal of Database Theory an Application, pp A Case Stuy of Applying SOM in Market Segmentation of Automobile Insurance Customers Vahi Golmah Department of Computer Engineering, Neyshabur Branch, Islamic Aza University, Neyshabur, Iran v.golmah@in.iut.ac.ir Abstract Over the last ecae, it has been observe that automobile insurer organizations are being taske with a new challenge characterize by increase competition, increase requirements of automobile insurance quality an an increasing emphasis on time-to-market. Furthermore, knowlege regaring what customers think, what they want, an how to serve them is quite useful for insurance organization wishing to generate suitable strategies in competitive markets. Keywors: Data mining, Self Organization Maps(SOM),Customer Relationship Management(CRM), Customer segmentation, Automobile insurance 1. Introuction Nowaays, how to make marketing strategies has been an essential issue for automobile insurance inustry [1]. Automobile insurance aims at covering ifferent types of claim incurre as a result of traffic accients which after joining Worl Trae Organization, insurance inustry has rapi growth. Over the last ecae, it has been observe that automobile insurer organizations are being taske with a new challenge characterize by increase competition, increase requirements of automobile insurance quality an an increasing emphasis on time-to-market. Furthermore, knowlege regaring what customers think, what they want, an how to serve them is quite useful for insurance organization wishing to generate suitable strategies in competitive markets. Owing to isparate esires, interests, an nees, gaining a comprehensive unerstaning of customers is ifficult. Since an organization cannot normally serve all customers in a market [2]. The iversity of customers' preferences an nees is a challenge for insurer organization to manage their customers through proviing a variety of attractive, personalize, an satisfactory service [3]. Therefore customer segmentation can improve archiving to marketing purposes. Customer segmentation ivies customers into groups, with the members of each group having similar nees, characteristics, or behaviors. Segmentation also represents the key element of customer ientification in customer relationship management.after segmenting customers, organizations can then use further strategies such as customer attraction to maintain relationships with customers an gain more profit from them, one-toone marketing, an etecting an preicting changes in customer behaviors'. Book mae by this file is ILLEGAL. ISSN: IJDTA

2 International Journal of Database Theory an Application Accoringly, this paper investigates the following research issues in Iran s automobile insurance business: What exactly are the customers' "nees" an "wants" for automobile insurance? How can help insurer to personalize its marketing campaigns? What woul be characteristics of new evelope insurance service to increase profit accoring to the knowlege of customers? Can the knowlege of customers be transforme into recommener systems or replenishment systems? Data mining tools are a popular means of analyzing to obtain a eeper unerstaning of each customer s behaviours. The knowlege extracte from ata mining results is illustrate as knowlege patterns an rules in orer to propose suggestions an solutions to automobile insurer. In the last ecae Self Organizing Maps (SOMs) have become a valuable ata mining tool for extracting the knowlege patterns an rules with process/reaction monitoring an unsupervise classification purposes [4-8]. In this paper, an attempt was mae to apply unsupervise Self Organizing Map for classification of customers of insurance services base on their behaviours an nees. The rest of this paper is organize as follows. The backgroun of the ata mining approach incluing clustering analysis, customer segmentation an Self Organization Map (SOM) process is reviewe in Section 2. Section 3 introuces our system framework an implementation of customer segmentation to analysis of customers an Section 6 presents a brief conclusion. 2. Research Methoology A competitive learning approach is propose for one-to-one marketing, an etecting an preicting changes in customer behaviors'. The iea behin this approach is to provie mechanisms for improving customer relationship management for automobile insurance organization through proviing a one-to-one marketing an variety of attractive, personalize, an satisfactory service base on their nees an behaviours. To analysis of evelope oneto-one marketing, it is necessary to give a brief review of ata mining, SOM, customer segmentation an their framework Data mining Due to the information technology improvement an the growth of internet, enterprises are able to collect an to store huge amount of ata. These massive atabases often contain a wealth of important ata that traitional methos of analysis fail to transform into relevant knowlege. Specifically, meaningful knowlege is often hien an unexpecte, an hypothesis riven methos, such as on-line analytical processing (OLAP) an most statistical methos, will generally fail to uncover such knowlege. Inuctive methos, which learn irectly from the ata without an a priori hypothesis, must therefore be use to uncover hien patterns an knowlege [9]. Book mae by this file is ILLEGAL. 26

3 International Journal of Database Theory an Application Figure 1. The phases of a ata mining process The explosive growth in atabases has enforce acaemia an inustry to use of ata mining techniques for extracting frequent structural patterns that may convey important information. Therefore, ata mining techniques has become an increasingly popular area for extracting information from the atabase in ifferent areas ue to its flexibility of working on any kin of atabases an also ue to the surprising results [7, 10]. Data mining is an interisciplinary fiel that combines artificial intelligence, atabase management, ata visualization, machine learning, mathematic algorithms, an statistics [11]. The life cycle of a ata mining project consists of six phases. Figure 1 shows the phases of a ata mining process. The sequence of the phases is not rigi. Moving back an forth between ifferent phases is always require. It epens on the outcome of each phase which phase or which particular task of a phase, has to be performe next. The arrows inicate the most important an frequent epenencies between phases. We provie a brief escription of these six phases in the following sections Business unerstaning This initial phase focuses on unerstaning the project objectives an requirements from a business perspective, then converting this knowlege into a ata mining problem efinition an a preliminary plan esigne to achieve the objectives Data unerstaning Book mae by this file is ILLEGAL. The ata unerstaning phase starts with an initial ata collection an procees with activities in orer to get familiar with the ata, to ientify ata quality problems, to iscover first insights into the ata or to etect interesting subsets to form hypotheses for hien information Data preparation The ata preparation phase covers all activities to construct the final ataset (ata that will be fe into the moeling tool(s)) from the initial raw ata. Data preparation tasks are likely to 27

4 International Journal of Database Theory an Application be performe multiple times an not in any prescribe orer. Basic tasks in ata preparation phase are as follows [5]: Data table an recor: The ata sources are first locate, accesse, an integrate. Next, selecte ata is put into a tabular format in which instances an variables take place in rows an columns, respectively. Data cleaning involves techniques for filling in missing values, smoothing out noise, hanling outliers, etecting an removing reunant ata. Data integration an transformation: Sometimes it is useful to transform the ata into a new format in orer to extract aitional information. It is useful to be able to summarize a large ata set of ata an present it at a high conceptual level. Dates are a goo example of ata that you may want to hanle in special ways. Any ate or time can be represente as the number of ays or secons since a fixe point in time, allowing them to be mappe. In the ata matrix, the month of the year is use instea of ate for etecting seasonal knowlege [12]. Data reuction an projection: This inclues fining useful features to represent the ata (epening on the goal of the task) an using imensionality reuction, feature iscretization, an feature extraction (or transformation) methos. Application of the principles of ata compression can play an important role in ata reuction an is a possible area of future evelopment, particularly in the area of knowlege iscovery from multimeia ata set [13]. Discretisation: This is a form of ata reuction, reuces the number of levels of an attribute by collecting an replacing low-level concepts with high-level concepts. real worl ata ten to be irty, incomplete, an inconsistent. Data preprocessing techniques can improve the quality of the ata, thereby helping to improve the accuracy an efficiency of the subsequent mining process. Data preprocessing is therefore an important step in the knowlege iscovery process, since quality ecisions must be base on quality ata. Detecting ata anomalies, rectifying them early, an reucing the ata to be analyze can lea to huge pay-offs for ecision making [13] Moeling Moeling, in software engineering, is the process of creating a ata moel by making escriptions of formal ata moels, using ata moeling techniques [9]. In this phase, various moeling techniques are selecte an applie an their parameters are calibrate to optimal values. Typically, there are several techniques for the same ata mining problem type. Data mining techniques accoring to the types of knowlege mine (DM functionalities) can classify to clustering, association, classification, an so on for achieving escriptive/preictive ata mining tasks [5]. Some techniques have specific requirements on the form of ata. Therefore, stepping back to the ata preparation phase is often necessary. Data mining moeling is the critical part in eveloping business applications. The goal of moeling is to formulate business problems as ata mining tasks. Moeling technology can provie quantitative methos for the analysis of ata, to represent, or acquire expert knowlege, using inuctive logic programming, or algorithms, so that AI, cognitive science an other research fiels are affore broaer platforms for the evelopment of DMT [7]. Book mae by this file is ILLEGAL Evaluation At this stage in the project you have built a moel (or moels) that appear to have high quality from a ata analysis perspective. Before proceeing to final eployment of the moel, it is important to more thoroughly evaluate the moel an review the steps execute to construct the moel to be certain it properly achieves the business objectives. A key objective 28

5 International Journal of Database Theory an Application is to etermine if there is some important business issue that has not been sufficiently consiere. At the en of this phase, a ecision on the use of the ata mining results shoul be reache. 2.2 Customer segmentation using of Self organization map Clustering an Segmentation is essentially aggregating customers into groups with similar characteristics such as emographic, geographic, or behavioral traits, an marketing to them as a group. Data mining techniques have been wiely applie to ifferent omains. As the transactions of an organization becomes much larger in size, ata mining techniques, particularly the clustering technique, can be use to ivie all customers into appropriate number of clusters base on some similarities in these customers. Facing the market with iverse emans, applying market segmentation strategy can increase the expecte returns. Much of marketing research focuses on examining how variables such as emographics an socioeconomic status can be use to preict ifferences in consumption an bran loyalty. Segmentation problem shoul be consiere as two ifferent situations known character parameters an unknown character parameters. Character parameters are known means segmentation analysis eals with customers who have transactional or behavioral recors store in the enterprise atabase an the analytic parameters are preefine an are erive from analyzer interests [9]. One of the possible clustering methos is competitive learning. Given the training set of objects, competitive learning fins an artificial object (representative) most similar to the objects of a certain cluster. A commonly use application of competitive learning is the Kohonen Self Organising Map, or SOM, escribe by Kohonen in 1982 [14]. SOM is a popular unsupervise neural network methoology to clustering for problem solving involving tasks such as clustering, visualization an abstraction an market screening. Compare to traitional clustering techniques such as the K-means algorithm, SOM has the following avantages. SOM is inspire by the cortex of the human brain, where information is represente in structures of 2D or 3D gris, the theory of which is motivate via observing the operation of the brain [6]. SOM is traine by an unsupervise competitive learning algorithm an can automatically etect strong features in large ata sets. While SOM maximizes the egree of similarity of patterns within a cluster an minimize the similarity of patterns belonging to ifferent clusters, it can prouce two-imensional arrangement of neurons from the multi-imensional space [6, 15]. Formally, SOM is a type of artificial neural network [11] with two fully interconnecte layers of neurons, the input layer an the output or Kohonen layer (it's shown in Figure 2). The first step of Kohonen learning is competition. Given the training vector on the network s input an weight vector for each neuron of the Kohonen layer, the neuron with the minimal (usually Eucliean) istance between the weight an input vectors is excite or selecte as the winner of the competition. The secon step is aaptation. The neurons of the Kohonen layer are organize in a one-, two-, or three-imensional lattice, reflecting its biological inspiration. A topological neighbor affecting function is efine on the Kohonen layer, assigning a egree of participation in the learning process to the neurons neighboring the winning neuron. In every learning step the weight vectors of the winning neuron an its neighbors are ajuste to move closer to the input training vector. The training algorithm propose by Kohonen for forming a feature map is state as follows [4, 6, 8]: Book mae by this file is ILLEGAL. Step (1) Initialization: Choose ranom values for the initial weights. 29

6 International Journal of Database Theory an Application Step (2) Winner Fining: Fin the winning neuron c at time t, using the minimum Eucliean istance criterion (1) where [ ] represents an input vector at time t, is the total number of neurons, an inicates the Eucliean norm. Step (3) Weights Upating: Ajust the weights of the winner an its neighbors, using the following rule: ( ) ( ) ( ) ( )[ ( ) ( )] (2) ( ) ( ) (3) ( ) Figure 2. The structure of self organization map (SOM) Where ( ) represents an input ata at time, ( ) is the topological neighborhoo function of the winner neuron c at time, ( ) is a positive constant calle "learningrate factor", an are the location vectors of noes c an, respectively. ( ) efines the with of the kernel. Both ( ) an ( ) will ecrease with time. It shoul be emphasize that the success of the map formation is critically epenent on the values of the main parameters (i.e., ( ) an ( )), the initial values of weight vectors, an the prespecifie number of iterations. Book mae by this file is ILLEGAL. In the case of a iscrete ata set an fixe neighborhoo kernel, the sum of the square-error of SOM can be efine as follows: (4) where is the number of training samples, an is the number of map units. Neighborhoo kernel is centere at unit, which is the best matching unit of input vector, an evaluate for unit. 30

7 International Journal of Database Theory an Application 3. Research Application To evelop operating strategies in competitive markets, the insurer organizations must first unerstan customer s characteristics an nees base on customer s expectations, an then ineffective strategies can be avoie, saving money an time. Thus, customer segmentation becomes important for insurance organization to recognize the customers, solving their problems, increasing their satisfaction, an enhancing their loyalty Iran automobile insurance at present To ate, there are a total of 23 insurance companies that all of them are local firms in active operation in Iran, since the first insurance organization was starte 78 years ago. In 2011, Iran was ranke 46th worlwie with total premium revenue of US$7 million accoring to the Swiss reinsurance company Sigma. This inicates that Iran s insurance market has expane to a significant position in the international insurance arena Data source Insurers are boun by the Private Data Protection Statute an are not permitte to isclose any information on their clients to other parties, which has restricte thorough research on existing automobile insurance policy purchases ata base on relate atabases. This stuy therefore collecte such ata by means of a questionnaire. Those subjects were automobile insurance consumers. Furthermore, the questionnaires are esigne to survey the subjects previous experiences with their purchases, an services renere to them by the insurers. Information collecte through the questionnaires is use to establish a atabase. The items in the questionnaires can be generally ivie into four parts: Part 1. Basic ata inclues seven questions: gener, age, eucation, occupation, resience, marital status an annual income. Part 2. Automobile characteristics inclue five questions: automobile type, automobile application, financial value of automobile, prouce year an county of its manufacturer. Part 3. Automobile insurance information inclues five questions: previous insurer organization, number of years which on t have any accient, the ays that past from perio of insurance an automobile insurer organization that wishes to purchase it as the future automobile insurance. Part 4. The importance of effective factors in selection of automobile insurance inclue five questions: insurance cost, iscount, paron for penalties cause to postponement of holing a new insurance, the satisfaction of previous insurer organization, the bran of insurer organization an native of insurer organization. Book mae by this file is ILLEGAL. Questionnaires were istribute among potential customers automobile insurance through the in Iran. A total of 2000 questionnaires were istribute, an 1307 were collecte. For this stuy, we get the ata use in this stuy from the collecte Questionnaires Data preparation The ata sources are first locate, accesse, an integrate. Next, selecte ata is put into a tabular format in which instances an variables take place in rows an columns, respectively. A labeling is use for a better unerstaning of results. By means of this labeling, a number is assigne to each questionnaire. 31

8 International Journal of Database Theory an Application Since the input to the ata mining moel affects the choice of a ata mining algorithm an the results, we attempte to remove pollute ata such as incorrectly coe input (e.g., typos) or inconsistent input (e.g., outliers or anomalous answers) from the atabase by filtering out the Excel file. Effective (1260) questionnaires were amasse other than 47 invali pieces containing omissions an incomplete answers. Effective collection ratio was 63%. The experimental ataset is ranomly ivie into two groups with 66% of the ata serving as a training set an the remaining 34% serving as a testing set SOM implementation In this section we use SOM toolbar in MATLAB software[16] to cluster customers into subgroups or segments. The computer use was a Pentium IV CPU at 2.5 GHz an 512 Mb of RAM. Before the training process begins, ata normalization is often performe. The linear transformation formula to [0, 1] is use: Where an represent the normalize an original ata; an represent the maximum values among the original ata. The etermination of the size of SOM is not an easy task, because a network with a greater number of cells woul have hinere the visualization of the labels in each neuron. In the same way, a smaller map than the one use by the authors woul cause many labels to be overlappe. It is recommene [17] that the number of neurons in the map training set an the length an with of the map shoul be proportional to the magnitue of the first eigenvalues obtaine by the ecomposition of the training set. The ratio calculate by the first two calculate eigenvalues in this case is Having that in min (an also the heuristic metho in SOM toolbar of MATLAB[16]) we starte the search for the optimal size of the map. After several trials, we chose a map with a size The use map ha plain bounary conitions, a hexagonal gri, Gaussian neighborhoo function, an linearly ecreasing learning rate. The SOMs were traine using the batch training algorithm in two phases: (1) rough training phase which laste 1000 iterations with an initial neighborhoo raius equal to 5, a final neighborhoo raius equal to 2, While a learning rate starts at 0.5 an en at 0.1, an (2) fine training phase which laste 500 iteration cycles, While a learning rate starts at 0.1 an en at 0.02, we o set the neighborhoo partitions starte at 2 an en with 0. After the training was finishe, the preiction abilities of the SOMs were examine using the statistical metrics an ata set consisting of 1260 customers. The evaluation base on the mean absolute error (MAE) an root mean square error (RMSE), which are wiely use for evaluating results of time-series forecasting. The MAE an RMSE are efine as follows: MAE= (6) Book mae by this file is ILLEGAL. RMSE= ( ) (7) Where an are the observe an the forecaste rice yiels at the time. By using of Eqs. (6) an (7) an ata of new cutomers, MAE an RMSE calculate as an , respectively. These low amounts for MAE an RMSE can emonstrate the appropriateness an precise of our moeling an forecasting. (5) 32

9 International Journal of Database Theory an Application 3.5. Customer segmentation It has been state in Section 2.2 that SOM map is a valuable tool to group (aggregate) an classify (isaggregate) customers base on their expectations an nees. This section explains how to use results of SOM to classify the customers an extracting the knowlege patterns an rules. U-matrix 5.27 Variable Variable2 Variable3 Variable Variable5 Variable6 Variable7 Variable8 Variable Variable10 Variable11 Variable12 Variable13 Variable Variable15 Variable16 Variable17 Variable18 Variable Variable20 Variable SOM 01-Feb-2013 Figure 3. The insurance customer ata set in SOM Figure 3 shows The insurance customer ata set in SOM base on each variable an its U-matrix. A U-matrix is a map of istances between each of neurons an all its neighbors an slices are the two-imensional combinations of the SOM weights. With the U-matrix map, the easiest way to examine whether there are istinct clusters present in a given ata set is to utilize the istance matrix. This matrix represents the istances between each neuron an all its neighboring ones, an is able to reveal the local cluster structure of the map. The further the istance is, the higher the ifference between them will be, which is resulte in a higher similarity. The istance between two neurons is represente by a color-coing scheme. Different colors are use to represent ifferent istances, such as re stans for the nearest neighborhoo istance, an grey stans for the farthest (Figure 4). Book mae by this file is ILLEGAL. Figure 4. U-matrix of customer ata set in SOM 33

10 International Journal of Database Theory an Application The 1307 customers of automobile insurance can be segmente to 4 follows clusters base on U-matrix that it is shown in Figure. a) Customers that cost is important for them: These customers have low annual income an financial value of automobile. They selom have any occupation an the insurance cost is an important factor to select insurance organization in this cluster (cluster 3). Cluster 3 has the least number of 312 (312/1260=0.248%) customers. They have the highest churn rating an it oesn t ensure organization/ customer relationship is maintaine in the long-term. An implication for this extracte phenomenon is that insurance organizations can have greater contribution of company s revenue from these customers by sening themselves iscount scheule to them. b) Customers that quality of services is important for them: Base on the information from U-matrix, Clusters 4 has the lowest number of customers. It has 120 (120/1260=9.5%) customers, while the highest monetary rating, 0.440, is gaine from its customers since their financial value of automobile is high. They choice an insurer organization as their automobile insurance which has a popular bran with high quality in evelope services. Moreover, they are loyal both in attitue an behavior in the automobile insurance. Loyal customers are profitable as they woul contribute positively to the success of their insurance experience. Therefore, the insurer organizations shoul focus on the customers of this cluster by increasing customer loyalty through proviing customize services to encourage their future purchase. c) Customers that style of payment is important for them: For Cluster 1, the average of eucation is above the total average. Besies, their occupation is as that they on t like pay all of costs in one time an prefer to pay in several installments. Pertaining to the marketing strategies for customers in Cluster 1, the insurer organizations shoul create payment convenience for this cluster which inclues special groups such as teachers, nurses an etc. ) Customers that select insurer organization base on family an frienly relation: Cluster 2 is a special group of patients for insurer organization. Cluster 2 has the most number of customers among clusters (438/1260=34.76%). They have ifferent attributes. They have the highest loyalty rating an it ensures organization/ customer relationship is maintaine in the long-term an the most revenues for the insurer organizations arise from them. In this regar, the customers of Cluster 2 can attract new customers. Since accoring to the so-calle rule, a ramatic business improvement is often achieve by ientifying the 20% of core customers an by maximizing the attention applie to them, they will attract the 80% of customers. Book mae by this file is ILLEGAL. Therefore, satisfying existing customers nees an buil close relationships with them will be very imperative. Pertaining to the marketing strategies for customers in Cluster 2, the insurer organizations shoul enhance customer relationship management to keep in touch with the customers an fin ways to meet their eman by attracting them to visit more often. For example, special attention an treatment is particularly vital to postgrauate customers when meet them an online services to simple relation to them. 34

11 International Journal of Database Theory an Application Figure 5. Graph of clusters obtaine with U-matrix To evaluate the preiction abilities of the SOMs we use the statistical metrics MAE an RMSE base on equations (6) an (7). The result for the training an testing ata sets using of SOM is shown in Table 1. Table 1. The statistical metrics MAE an RMSE for the training an forecast using SOM on ata set Training Testing MAE RMSE MAE RMSE The low MAE an RMSE in Table 1 for training an testing ata sets shows high performance SOM for clustering of insurance information ata set use in this research. 4. Conclusion The complex organizational structure of most of the companies that write a significant volume of automobile insurance makes it infeasible for a ecision maker to cope with all unerwriting ecisions by using of traitional analysis methos. Hence, it is necessary to esign some set of strategies an guielines to ensure a maximum egree of consistency among customers. Knowing customers one by one in some business is not commoious, so using customer segmentation metho coul be useful in this situation. Customer segmentation base on their nees, characteristics, or behaviors. Segmentation also represents the key element of customer ientification in customer relationship management. After segmenting customers, organizations can then use further strategies such as customer attraction to maintain relationships with customers an gain more profit from them, one-to-one marketing, an etecting an preicting changes in customer behaviors'. Accoringly, this paper investigates the CRM research issues in Iran s automobile insurance business by applying of unsupervise self-organizing maps for classification of customers of insurance services base on their behaviours an nees. To evaluate the preiction abilities of our moel base on SOMs we use the statistical metrics MAE an RMSE. Moreover, the secon criterion for moel valiity in this stuy was the views of management strategists. The management strategists of the insurance organizations on which the case stuy was conucte have sai that they have foun the results obtaine from the suggeste moel to be meaningful an useful. Thus, as of August 2006, in view of the Book mae by this file is ILLEGAL. 35

12 International Journal of Database Theory an Application strategies followe in the last year the strategists claim that the strategies etermine by the moel in our stuy in May 2005 are inee accepte as the best solutions. Although we focus on automobile insurance, our approach may help managers in a variety of contexts to assess other insurances. Future research may research to life insurance. In aition, fuzzy numbers can be introuce in SOM methos to more effectively analyze cases having greater. Acknowlegments The authors woul like to thank the Eitor-in-Chief, associate eitor, an unknown referees for their useful comments an suggestions, which were very helpful in improving this stuy. References [1] S. Anwar an M. Zheng, "Competitive insurance market in the presence of ambiguity", Insurance: Mathematics an Economics; (2012). [2] L. Bermuez an D. Karlis, "Bayesian multivariate Poisson moels for insurance ratemaking", Insurance: Mathematics an Economics, vol. 48, (2011), pp [3] M. Eling an M. Luhnen, "Efficiency in the international insurance inustry: A cross-country comparison", Journal of Banking & Finance, vol. 34, (2010), pp [4] M. H. Ghaseminezha an A. Karami, "A novel self-organizing map (SOM) neural network for iscrete groups of ata clustering", Applie Soft Computing, vol. 11, (2011), pp [5] G. Koksal, Batmaz an M. C. Testik, "A review of ata mining applications for quality improvement in manufacturing inustry", Expert Systems with Applications, vol. 38, (2011), pp [6] P. Klement an V. Snášel, "Using SOM in the performance monitoring of the emergency call-taking system", Simulation Moelling Practice an Theory, vol. 19, (2011), pp [7] S. H. Liao, P. H. Chu an P. Y. Hsiao, "Data mining techniques an applications: A ecae review from 2000 to 2011", Expert Systems with Applications, vol. 39, (2012), pp [8] S. L. Shieh an I. E. Liao, "A new approach for ata clustering an visualization using self-organizing maps", Expert Systems with Applications, vol. 39, (2012), pp [9] J. L. Seng an T. C. Chen, "An analytic approach to select ata mining for business ecision", Expert Systems with Applications, vol. 37, (2010), pp [10] N. Aggarwal, A. Kumar, H. Khatter an V. Aggarwal, "Analysis the effect of ata mining techniques on atabase", Avances in Engineering Software, vol. 47, (2012), pp [11] H. H. Tsai, "Global ata mining: An empirical stuy of current trens, future forecasts an technology iffusions", Expert Systems with Applications, vol. 39, (2012), pp [12] C. Ciflikli an E. Kahya-Ozyirmiokuz, "Implementing a ata mining solution for enhancing carpet manufacturing prouctivity", Knowlege-Base Systems, vol. 23, (2010), pp [13] F. Gürbüz, L. Özbakir an H. Yapici, "Data mining an preprocessing application on component reports of an airline company in Turkey", Expert Systems with Applications, vol. 38, (2011), pp [14] T. Kohonen, "Self-Organizing Maps", Berlin: Springer-Verlag; (1995). [15] J. T. Wei, S. Y. Lin, C. C. Weng an H. H. Wu, "A case stuy of applying LRFM moel in market segmentation of a chilren's ental clinic", Expert Systems with Applications; (2012). [16] J. H. E. Alhoniemi, J. Parviainen an J. Vesanto, "Som toolbox for matlab", (1999). [17] I. Kuzmanovskia, M. Trpkovskaa an B. Soptrajanov, "Optimization of supervise self-organizing maps with genetic algorithms for classification of urinary calculi", Journal of Molecular Structure, vol , (2005), pp [18] R. S. Wu an P. H. Chou, "Customer segmentation of multiple category ata in e-commerce using a softclustering approach", Electronic Commerce Research an Applications, vol. 10, (2011), pp Book mae by this file is ILLEGAL. 36