Avalable onlne www.jocpr.com Journal of Chemcal and Pharmaceutcal Research, 204, 6(6):698-703 Research Artcle ISSN : 0975-7384 CODEN(USA) : JCPRC5 Customer segmentaton, return and rsk management: An emprcal analyss based on BP neural network L Zhou and Qng-y Chen 2* School of Management, Mnzu Unversty of Chna, Bejng, Chna 2 School of Busness, Renmn Unversty of Chna, Bejng, Chna ABSTRACT In an artfcal neural network, smple artfcal nodes called "neurons" are connected together to form a network whch mmcs a bologcal neural network. BP neural network has been wdely used n tranng artfcal neural network. In ths paper, by usng BP Neural network and data mnng method we fnd out the nfluence factor of optmal compensaton rate n nsurance companes. Accordng to the data analyss, the nsurance company should make correspondng strateges: frst, make a relatve low settlement rato to the customers who have drvng experence n -5 years, especally n -2 years; second, make a hgher settlement rato to the customers that have drvng experence more than 5 years, the fnal purpose s to keep the exstng customers, and be able to attract some potental customers. Key words: Data Mnng method, Customer Relatonshp Management, BP Neural network, Fnancal Insttutons INTRODUCTION At the begnnng of the 990's, a new busness model appeared n the Unted States and called Customer Relatonshp Management, CRM s a new management mechansm that ams to mprove the relatonshp between enterprses and customers []. CRM provdes comprehend-sve, personalzed customer nformaton to sales and servce personnel, t also strengthen the trackng servce, nformaton analyss capabltes, enablng them to buld and mantan the one-to-one relatonshp" between customers and enterprse [2], so that enterprses can provde more effcent servce, mprove customer satsfacton, attract and retan more customers. Customer segmentaton as one of the core concept of customer relatonshp management, s refers to the enterprse whch has explct strateges, busness models and specfc market, provdng products, servces accordng to customer attrbutes, behavor, preferences, needs and values and other related factors [3]. Wth the extensve applcaton of management nformaton system, enterprse wll accumulate more and more data, the tradtonal customer segmentaton method can not deal wth t [4]. Data Mnng s a data analyss technology based on artfcal ntellgence, ts man functon s to dscovery useful nformaton from large amounts of data, whch wll help enterprses to manage customer resources effectvely. Gannet Group frst put forward the concept of CRM (customer relatonshp management), t consder the customer relatonshp management s to provde a full range management perspectve, gve more customer communcaton sklls to enterprses, and fnally maxmze customer return rate [5]. Accordng to IDC's (Internatonal Data Corporaton) research, the mean ROI( return on nvestment)of company whch use CRM can reach 400%,more than 90% enterprses ROI exceeded 40%, 50% enterprses ROI exceeded 60%, 25% enterprses ROI over 600%.Data mnng n customer relatonshp management applcatons can contrbute sgnfcantly to the bottom lne, as the development of nformaton technology, customer nformaton ncrease fast n recent years. So that, data mnng s a useful method to dscover extract nformaton n large data sets. As customer segmentaton s the core functon n CRM, data mnng can also be used to automatcally dscover the segments or groups wthn a customer data set [6]. 698
Usama Fayyad& Gregory P. Shapro (996) ponted out that data mnng and knowledge dscovery n databases have been attractng a sgnfcant amount of research and ndustry, nvolves sx common classes of tasks as: anomaly detecton, dependency modelng, clusterng, classfcaton, regresson and summarzaton[7].there are also many researchers focused on evaluatng the data mnng applcatons n fnancal ndustry. Georgos Sarantopoulos (2003) presented a real-world applcaton of a data mnng approach to credt scorng, and ponted out that the decson tree showed a great mprovement n performance compared to the current manual decsons [8]. Wen-bng Xao & Qan Zhao (2006) nvestgated the performance of varous credt scorng models and the correspondng credt rsk cost for three real-lfe credt scorng data sets[9]. Xaohua Hu (2005) presented a data mnng approach for analyzng retalng bank customer attrton, and demonstrated effectveness and effcency by applyng data mnng technology n retalng bank [0]. In these studes, however, dd not fully play the bass of customer segmentaton functon, some s only a smple object wth the customers, also there are few artcles combne customer segmentaton and data mnng technology. So that, I use data mnng technology n ths paper to help nsurance company mprove CRM by makng effectve market segmentaton and optmal prce. CRM mantans and manages customer relatonshp through the four stages as: dentfcaton of customer, analyss of the dfference, bengn contact and the customzed servce. Customer relatonshp management requres "customer satsfacton frst " as the core phlosophy, so customer servce s the core busness n CRM. EXPERIMENTAL SECTION 2. Artfcal Neural Network The nspraton for neural networks came from central nervous systems and has been wdely used n nformaton technology. In an artfcal neural network, smple artfcal nodes called "neurons" are connected together to form a network whch mmcs a bologcal neural network. BP neural network has been wdely used n tranng artfcal neural network (ANN). The BP neural network has three layers, shown as fgure.the frst layer has nput neurons, whch send data va synapses to the second layer of neurons, and then va more synapses to the thrd layer of output neurons. Input Hdden Output Fgure. BP neural network The key to establsh a neural network model s how to select the approprate connecton weght matrx, ths task s realzed by network tranng, and the basc prncple s as follows. Input I (n) to the correspondng nput layer, then output layer wll get the output as Y (n), then let D (n) represents the correspondng output that gven by the tranng sample, so the error sgnal at the j output neurons can be expressed as: e ( n) d ( n) y ( n) = () Assume the samples number for tranng s N, then the square error to the total tranng sample would be: E r = N N n= E ( n) (2) 699
So the goal s get a set of weghts W to make the functon Er acheve the mnmum value, as: Mn N N 2 n= = c 2 ( d ( n) y ( n) ) (3) 2.2. Structure Model In customer segmentaton structure model, the basc work of data mnng method s through analyzng the known data and then summed up a forecast model. The data can be ether hstorcal data or exogenous data; exogenous data can be obtaned by expermental method and research method. Data mnng method of customer segmentaton should have the followng abltes: a) The dynamc behavor descrpton b) The relablty of the data c) The nose s resstance and tme-varyng d) Synthess of varous mnng methods The mplementaton process of customer segmentaton s based on sample learnng method. When decde the marketng strategy of customer relatonshp management, managers often use some descrpton of customer characters, such as hgh-ncome customers, low-ncome customers, trendy customer, conservatve clents, hgh rsk customer, the man task of customer segmentaton s to ensure the correspondng relatons between t these concepts and the correspondng customers. Customer data contans several dscrete customer attrbutes and contnuous customer attrbutes, each customer attrbute as a dmenson, every customer as a bt of space, the enterprse customer database of all customers can consttute a multdmensonal space, called the attrbute space of customers. Assume A={A, A 2, A m } Is A set of propertes that descrbe customer characterstcs and behavor, these propertes can be ether contnuous attrbutes or dscrete attrbutes, these attrbutes form the m-dmensonal space A, each customer value determnes the poston n space A. We assume c C, the value of c n attrbute A s c[a ].we use g as a abstract concept that descrpt customer, f(g) s the set of customer. For a set of concepts{g, g 2, g k }, f {f(g ),f(g 2 ), f(g k )} s k mutually dsjont sets at any moment, and C= f(g ) f(g 2 ) f(g k ),then we can get G = g g 2 g k. Accordng to c C, f c f(g j ) and j k, t can be recorded as ϕ(c, G )=g j. In the concept of customer value dmensons, there are three concepts as valued customer, potental value customer, no value customers, these three concepts can summarze all customers. We assume B G, R B C C s a bnary relaton of C, and C C s a Cartesan product of C, the formula of R B s: {( c c ) c, c C, b B, ϕ [ c, b = ( c b) ]} R B =,, 2 2 ϕ 2 (4) In ths formula, R B s an equvalence relaton, customers can be dvded nto several class of the space, and each equvalence class s called a concept class. Customer segmentaton s the process that bult the mappng relatonshp of customer attrbute space A m and concept space G n as: A m G n. The sampler learnng method s manly through data mnng process about the customer data, ths data mnng process was shown as fgure 2. We assume B={G, G 2, G k }, so that L={L, L 2, L k }, c C. c s the customer set that concept class has been known, customer segmentaton should fellow two steps as: a) Defne a mappng as p: C L, whch makes c C, f c L, so p(c)=l ; b) c C, determne the class by calculate p(c). 700
CRM database Tranng data set Classfcaton tranng algorthm Categorcal data Classfcaton model Result Fgure 2. The mplementaton model of customer segmentaton Node 0 % 00 Perod.5 >.5 Node Node 2 % 8.5 % 9.5 Use Use Enterprse Home Government Tax Bus Node 3 Node 4 Node 5 Node 6 Node 7 % % 2.5 % 5 % 2 % 89.5 Fgure 3. Decson Tree 2.3. Emprcal Analyss In ths paper, I use data mnng method to fnd out the nfluence factor of settlement n nsurance companes. The nsurance company pad most attenton to the amount of settlement, how to reduce the amount of settlement and how to formulate the correspondng settlement rate s mportant for nsurance company survve. The factors whch nfluence the amount of settlement, s the prmary soluton to the nsurance company. In fact, the relatonshp between these factors s very complex, so we use data mnng technology and decson tree method to make a further research on the nfluence factors of the amount of nsurance settlement. 70
The data s collected from Chna nsurance company; t ncludes the settlement records of car owners. Frst, I put the data nto Clementne, and then make standardzaton processng of the data by usng data dscretzaton and data generalzaton to remove out the redundant nformaton. Then, I use feature selecton to choose the man factors that affect the amount of settlement and analyze the customer records. The result shows that: the tme perod of drvng experence and the purpose of usng a car s the man factors affect the amount of nsurance settlement. Usng the Clementne decson tree model and enhanced ID3 algorthm, we can get the decson tree as shown n fgure 3. From the decson tree model, we can get that: frst, the average settlement amount s 434 RMB, so that we should ncrease the consumer rato that gets average settlement less than 434 and make a hgher nsurance prce for those who get settlement amount more than 434; second, we can fnd that the person who has the drvng experence less than.5 year needs a average settlement as 5040 RMB, ths part of the customer has brought a heavy burden to the company. Therefore, the company should gve ths part of customers a relatvely low settlement rate or ncrease the nsurance prce to these customers; thrd, we can fnd that usng cars for enterprse, home, government and tax often has a hgh settlement than 434 RMB, however, usng cars as bus often has a especally low settlement, so that the nsurance company should pay more attenton to those bus nsurance. 2.4. Influence Factor Analyss To sum up, we can know the amount of settlement s manly nfluenced by the drvng experence and the purpose of usng, and then I wll make a further analyss on these nfluence factors. I use the Hstogram functon n Clementne to analyze the nfluence factor as drvng experence, the result were shown n fgure 4 and fgure5. From fgure 4, we can fnd that settlement events occurred manly n those customers who have the drvng experence between -5 years, but less occurred n customers who has drvng experence more than 5 years. From fgure 5, we can fnd that the customers who have a drvng experence between -2 years often have a hgher settlement amount. Fgure 4.tme perod of drvng experence Fgure 5. Settlement Amount Accordng to the data analyss, the nsurance company should make correspondng strateges: frst, make a relatve low settlement rato to the customers who have drvng experence n -5 years, especally n -2 years; second, make 702
a hgher settlement rato to the customers that have drvng experence more than 5 years, the fnal purpose s to keep the exstng customers, and be able to attract some potental customers. CONCLUSION Ths paper frst ntroduced customer relatonshp management theory and customer segmentaton, and then used data mnng method to fnd out the nfluence factor of settlement. Frst the essay puts forward the prncple of customer segmentaton on the base of exstng problems of customer segmentaton n customer relatonshp management. Then from the logcal model of customer segment, and puts forward the model of data - functon - methods of customer segmentaton. In the case study, I used the decson-makng tree and the vsual functon of the Clementne to analyze the nfluence factor of nsurance settlement. The results showed that the tme perod of drvng experence and the purpose of usng a car s the man factors affect the amount of nsurance settlement. Then I used that data to make a further analyss, and fnd out that the consumers who have the drvng experence between -2 years have a hgher rsk to get a car accdent, these consumers got an average settlement as 5040 RMB, and t s much hgher than the average settlement of total consumers. So that the nsurance company should make a relatve low settlement rato to these customers, and make effectve customer segmentaton based on data mnng method. REFERENCES [] R Jayant;B Vsha, Internatonal Journal of Informaton and Communcaton Technology,200, 2, 89-99 [2] T Yamaguch, Artfcal Lfe and Robotcs, 2009,4, 45-53 [3] A Tuzhln,Data Mnng and Knowledge Dscovery, 202, 424-430 [4] GE Fruchter, SP Sgué, Journal of Optmzaton Theory and Applcatons,2009,42, 32-38 [5] L Yu; S Wang;KK. La, Fronters of Computer Scence n Chna,200, 4, 02-06 [6]G Monka;V Rajan, Internatonal Journal of Computer Scence Issues, 202,9, 22-29 [7]F Usama;PS Gregory;S Padhrac, AI Magazne,996, 7,96-20 [8]G Sarantopoulos,Operatonal Research, 2003, 3,56-66 [9]W Xao;Q Zhao;Q Fe,Journal of Systems Scence and Systems Engneerng,2006,5, 42-50 [0]X Hu,Appled Intellgence, 2003, 22, 2-23 703