A Change Detection Model for Credit Card Usage Behavior

Transcription

1 Proceedngs of he 5h WSEAS In. Conf. on COMPUTATIONAL INTELLIGENCE, MAN-MACHINE SYSTEMS AND CYBERNETICS, Vence, Ialy, November 20-22, A Change Deecon Model for Cred Card Usage Behavor CHIEH-YUAN TSAI *, JING-CHUNG WANG, CHIH-JUNG CHEN Deparmen of Indusral Engneerng and Managemen, Yuan-Ze Unversy 135 Yuan-Tung Rd., Chung-L, Taoyuan TAIWAN, R.O.C. Absrac: In recen year, cred card has been one of he mos aracve fnancal producs all over he world. The magnfcen ncrease n cred card marke leads card ssuers pu more aenons on how o undersand her cusomers. Ths research proposes a deecon model o denfy usage behavor change paerns of cred card cusomers n wo me perods. In he model, cusomer profles and purchase ransacons of wo me perods are rereved from card ssuer daabases. Then, a usage behavor rule se for each me perod s generaed usng Apror assocaon rule algorhm. Fnally, sgnfcan usage behavor change paerns are denfed hrough rule se comparson usng defned smlary and dfference measures. The proposed model has been successfully mplemened usng real cred card daa provded by a commercal bank n Tawan. Several markeng sraeges are suggesed accordng he analyss fndng. Wh he proposed model, card ssuers can deec crcal usage behavor changes and allocae her lmed resource o esablsh suable markeng sraegy for her cusomers. Key-words: Daa Mnng, Change Deecon, Assocaon Rule, Cred Card Usage, and Cusomer Relaonshp Managemen. 1 Inroducon Accordng o he Nlson Repor, purchases of goods and servces blled o cred card ssued n he U.S. reached $1.472 rllon n Cred card and deb card purchase volume n 2002 was $1.852 rllon, up 9.8% from 2001 [1]. Vrually, cred cards have be used everywhere from vacaons, busness rps, grocery sores, and resaurans. Wheher you call hem bank cards, gas cards, real cards, ravel and eneranmen cards, or smply plasc cards, here s no doub ha he cards have revoluonzed he busness model and become an essenal elemen of our daly lfe. One sraegy o ncrease he prof of card ssuers s o nensfy her compeon n marke hrough ncreasng sasfacon, reenon, and loyaly of cusomers [2] [3] [4]. Tha s, card ssuers emphasze on undersandng wha her cusomers purchase, when hey use he card, and how ofen hey consume. When he usage behavor nformaon s avalable, he card ssuers can encourage cusomers use her cards more frequenly hrough offerng beer producs and servces. Daa mnng s he echnque o dscover meanngful paerns (rules) and consruc models from large daabases. Much of exsng daa mnng research has focused on devsng echnques o buld accurae models and o dscover rules. Relavely lle aenon has been made o mnng changes n daabases colleced over me [4] [5] [6]. Emergng paern mnng s he process o dscover sgnfcan changes or dfferences from one daabase o anoher [7]. Emergng paern capures emergng rends n me samped daabase. Anoher relaed research rend s subecve neresngness mnng. Ineresngness mnng s o fnd unexpeced rules wh respec o he user s exsng knowledge. Unexpeced changes compare each newly generaed rule wh each exsng rule o fnd degree of dfference [8]. Lu e al. [9] proposed a DM- II (Daa Mnng-Inegraon and Ineresngness) sysem whch has classfcaon and assocaon rule mnng asks o help users perform neresngness analyss of he rules. Is analyss compares each newly generaed rule wh each exsng rule o fnd degree of dfference, whch s useful and mporan for reallfe daa mnng applcaons. Han e al. [10] presened several algorhms for effcen mnng of paral perodc paerns, by explorng some neresng properes relaed o paral perodcy. The algorhms show ha mnng paral perodcy needs only wo scans over he me seres daabase o make effcen n mnng long perodc paerns. Ths research proposes a deecon model o denfy cred card usage behavor changes n wo me perods. Three change paerns (emergng paern, unexpeced change, added/pershed rule) are

2 Proceedngs of he 5h WSEAS In. Conf. on COMPUTATIONAL INTELLIGENCE, MAN-MACHINE SYSTEMS AND CYBERNETICS, Vence, Ialy, November 20-22, defned and deeced usng desgnaed smlary and dfference measures. The res of hs paper s organzed as follows. Secon 2 nroduces he framework of he proposed change deecon model. Secon 3 provdes an mplemenaon case usng a real cred card daabase provded by a commercal bank n Tawan o demonsrae he benef of he proposed model. A summary and fuure works are concluded n Secon 4. 2 A behavor change deecon model The framework of he proposed change deecon model for cred card usage behavor s dvded no hree sages as shown n Fgure 1. Daase a m e Rulese for m e $ A C ard Issuer D aa exracon and preprocessng A ssocaon R ules G eneraon Calculae Sm lares and D fferences Evaluae Degree of Change Sgnfcan Changed Rulese M anageral Suggeson Daase a m e +k Rulese for m e +k Fg. 1: The framework of he proposed sysem. The frs sage The second sage 2.1 Daa exracon and preprocessng The frs sage s o rereve he cusomer profle and her ransacon daa from daabases. The ransacon daa s dvded no wo daases accordng o he ransacon dae we wan o sudy. To undersand he usage behavor of cred card cusomers, he daa relaed o cusomer profle and consumer behavor need o be exraced. Praccally, demographc arbues such as Age, Gender, Marrage, Educaon, Occupaon, Address, Cred Saus and are ofen used o descrbe a cusomer profle, whle he ransonal daa arbues such as Transacon Dae, Sore Address, merchan caegory codes (MCC), and Transacon Amoun are used o descrbe consumer behavor n cred card busness. Noes ha, connuous values n he daa such as Age, Income, and Transacon Amoun should be dscrezed o approprae caegorcal values o The hrd sage faclae he assocaon rules generaon n he nex sage. Le I = { 1, 2,, m} be a se of ems where each em s an arbue-value par. The par could be a cusomer profle or consumer behavor arbue wh respecve caegorcal value. Le he daase D = { 1, 2,, n} be a se of records and = { I 1, I2,, Ik} be a se of arbue-value par ems for a cusomer where I I. Snce hs research nends o deec he usage behavor change paerns n wo me perods, he daase D s separaed o daases D k and D + where D k and D + represen he daases sorng he records ransaced n me perod and +k respecvely. 2.2 Assocaon rules generaon The second sage s o generae rule ses from he wo daases usng assocaon rule echnque. k Each daase ( D or D + ) can be used o generae a rule se for s me perod ( or +k) usng assocaon rule echnque. A rule se s a collecon of assocaon paerns descrbng neresng causal relaonshp beween cusomer profle and consumer behavor. An assocaon rule s an mplcaon of he form LHS RHS where LHS, RHS I and LHS RHS = φ. The suppor for an assocaon rule LHS RHS s he percenage of reco rds n he daabase ha conans LHS RHS. If he suppor s hgh, he emse LHS RHS s worh o pu no dscusson. Noes ha a se of ems s referred o an emse. The confdence for an assocaon rule LHS RHS s he rao of he number of records ha conan LHS RHS o he number of records ha conan LHS. Therefore, he goal of he assocaon rule problem s o fnd ou all assocaon rules LHS RHS wh a mnmum suppor ( s mn ) and a mnmum confdence ( mn ) where he s mn and mn are specfed by a user. A well-known assocaon rule algorhm, called Apror algorhm [11], s ulzed n hs research. The algorhm uses a so-called large emse propery o reduce he compuaonal me. The propery saes ha any subse of a large emse mus be large. Afer he large emses have been found, s sraghforward o generae srong assocaon rules, where srong assocaon rules sasfy boh mnmum suppor and mnmum confdence. 2.3 Behavor change deecon

3 Proceedngs of he 5h WSEAS In. Conf. on COMPUTATIONAL INTELLIGENCE, MAN-MACHINE SYSTEMS AND CYBERNETICS, Vence, Ialy, November 20-22, In he hrd sage, he wo rule ses are compared usng smlary and dfference measures. Afer obanng he rule se R from me perod and k R + from me perod + k, he nex sage s o fnd ou he usage behavor changes beween he wo me perods. In he followng descrpon, r represens he h rule n k R and r + be he h rule k n R +. Sup ( r ) represens he suppor of r n me. In hs research, hree usage behavor change paerns are defned. Frs, r +k s consdered as an emergng paern respecve o r when boh LHS and RHS of r and r +k are he same bu here s a k sgnfcan dfference n Sup ( r ) and Sup + ( r ) [7]. Second, r +k s an unexpeced ch ange wh respec o r, f LHSs of r and r +k are smlar bu her RHSs are que dfferen [8]. Thrd, r +k s an added rule f all he LHSs and RHSs are que dfferen from any of r n R, and r s a pershed rule f all he LHSs and RHSs are que dfferen from any of r +k k n R + [12]. Wheher rules are smlar or que dfferen can be udged by a Rule Machng Threshold (RMT) [13]. The value s subecvely provded by a user. The relaon beween RMT and he hree change paerns s llusraed n Fgure 2. C om pleely D fferen Added/ Pershed Rule RM T value U nexpeced Change Compleely Sam e Emergng Paern Fg. 3: The hree usage behavor change paerns. For beer explanaon, some noaons and varables are defned: +k s s he smlary measure beween r and r where 0 s 1. +k s he dffe rence measure beween r and r where s he degree of arbue mach n he LHS + k par, where l = A max( X, X ). c s he. par degree of arbue mach n he RHS A s he number of he same arbues n he LHS pars of r and r +k. X s he number of arbues n he LHS pars of r. k X + s he number of elemens n he LHS pars n r +k. x k s he degree of value mach of he kh machng arbue n he LHS pars. y s he degree of value mach n he LHS pars. The change paern can be deeced hrough he followng seps [13]: 1. Frs, calculae he maxmum smlary value for each rule n me perod and +k. The maxmum smlary value of r s defned as s = max( s 1, s2,, s ) R and he maxmum smlary + k value of r +k s defned as s = max( s1, s2,, s ), where s s he smlary measure beween rule r and r +k. The descrpon of s s defned as s = x k A k A c y, f A 0, f A 0 = 0 If boh LHS and RHS s he same, he value of smlary wll be 1. Ohers wll have he value beween 0 and Second, for each rule r, calculae he dfference measures and modfed dfference measures beween r and r +k. The dfference measure can be defned as: xk k A y, f A 0, c = 1 = A y, = 0, = 1 f A c The modfed dfference measure can be defned as: k, 1, f max( s, ) = = s 1 = k 0, oherwse 3. Thrd, he usage behavor change paerns are udged usng he maxmum smlary values ( s and s ), dfference measures ( ), and modfed dfference measures ( ). If = 0 ( xk > 0or y > 0or > 0), hen he change k A ype s an emergng paern. If R > 0, RMT, hen he change ype s an unexpeced consequen change. If < 0, RMT, hen he change ype s an unexpeced condon change. If s < RMT, hen

4 Proceedngs of he 5h WSEAS In. Conf. on COMPUTATIONAL INTELLIGENCE, MAN-MACHINE SYSTEMS AND CYBERNETICS, Vence, Ialy, November 20-22, he change ype s an added rule. If s < RMT, hen he change ype s a pershed rule. Noes ha he number of changes could be large n mos cases. Therefore, only hose paerns whose degrees of change are no less han a mnmum degree of change specfed by users mn, called sgnfcan changes, need o be examned. The degree of change for each change paern can be shown n Table 1. Table 1: The change degree for dfferen change ypes. Type of changes Measures Em ergng Paen +k ( Sup ( r ) Sup ( r )) / Sup ( r ) Une xpeced change + k k Sup r ) / Sup + r ) Added rule Pershed rule 3 A case sudy ( ( 1 s ) Sup ( r + k ( 1 s ) Sup ( r The proposed framework s mplemened usng he daa provded by a maor cred card ssuer n Tawan. The frs me perod for hs sudy s year 2001 and he second perod s year There are oally 464,621 acve and nacve cred card ssued unl Daa preprocessng Afer a comprehensve dscusson wh he markeng managers of he bank, he daa selecon crera are se such as card ssued for more han 6 monhs, no delay paymen n he las wo monhs, he oal spendng greaer han 200,000 NT n he las 12 monhs and so on. In addon, welve crcal cusomer profle and consumer behavor arbues are denfed. They are Cusomer ID, Gender, Age, Marrage, Educaon, Occupaon, Locaon, Cred Saus, Yearly Usage Frequency, Yearly Spendng Amoun, Merchan Caegory Code, and Spendng Amoun. A seral of COBOL (common busness orened language) and JCL (ob conrol language) programs are coded o rereve cusomer profles and consumer behavor daa from he VSAM (vrual sorage access mehod) fles n OS/390 operaon sysem of an IBM 9121man frame compuer. There are 10,940 cusomers who sasfy he selecon crera n boh year 2001 and ,066 ransacon records are found n year 2001 and 467,504 ransacon records n year 2002 for hose cusomers. The dscrezaon s processed ) ) ( eravely based on he daa dsrbuon of each arbue. 3.2 Assocaon rules for he wo me perods Dependng on praccal need, a user can decde an approprae mnmum suppor value ( s mn ), mnmum confdence value ( mn ), and he larges emses lengh. If he mnmum suppor s hgh, he number of assocaon rules wll b e low. Alernavely, f he mnmum suppor s low, he number of assocaon rules wll be many. For nsance, n year 2001 here are 63,288 assocaon rules generaed when mnmum suppor s 1% and 6,312 when mnmum suppor s 9% (whle mnmum confdence s 70% and he larges emse s 4). I s found ha he number of rules decreases dramacally when he mnmum suppor ncreases. In he followng dscusson, he mnmum suppor s se as 1%, mnmum confdence as 70%, and he larges emse as 4. Wh he seng, we ge 63,288 assocaon rules n year 2001 and 56,571 assocaon rules n year Afer deleng he meanngless rules, 1,986 assocaon rules n year 2001 and 1,622 assocaon rules n year 2002 are lef for he followng change deecon analyss. 3.3 Change deecon analyss As descrbed n secon 2.3, he number of unexpeced changes rules and added/pershed rules s affeced by he RMT. For example, 33 unexpeced changes and 11,724 added/pershed rules are denfed f he RMT s se as 0.4, whle 6,323 unexpeced changes and 614 added/pershed rule are found f he RMT s 0.8. Noed ha he number of emergng paerns s no affeced by he RMT value. Due o he number of rules for he hree change ypes s large, we need o screen ou some changes ha mgh no be sgnfcan enough o pay aenon on. Tha s, he degree of change greaer han mnmum degree of change ( mn )wll be furher consdered. For example, f we se mn =0, hen 8,122 sgnfcan emergng paerns, 759 sgnfcan unexpeced changes, and 264 sgnfcan added/pershed rules wll be sgnfcan. Sgnfcan emergng paerns Sgnfcan emergng paerns are he rules ha appear n year 2001 and 2002 bu her suppor levels are que dfferen. There are 8,122 sgnfcan paerns found f we se mn =0. Three

5 Proceedngs of he 5h WSEAS In. Conf. on COMPUTATIONAL INTELLIGENCE, MAN-MACHINE SYSTEMS AND CYBERNETICS, Vence, Ialy, November 20-22, of hem are summarzed n Table 2. As shown n Table 2, he suppor levels of Paern 1 o Paern 3 ncrease from abou 1 % n 2001 o 45% n I s abou 40 mes ncrease for hose hree paerns. Wh hose paerns, we know ha male cusomers who lve n Tape, Tachung, or Kaoshung (op hree larges ces n Tawan) wh college or lower educaon level, dramacally ncrease her purchase n auo servce and gas saon from year 2001 o year We furher query he number of cusomers ha sasfy he hree paerns and fnd ha here are 3,211, 1,508, and 1,606 persons sasfyng he Paern 1 o 3 respecvely. If he bank wshes o promoe auo servce and gas saon program o s cusomers, up o 42.1% of malng fee can be saved by us sendng he mals o he neresng group nsead of all cusomers Table 2: Sgnfcan emergng paerns. Year Year Paern Year 2001 (or 2002) Sup Sup, 1 Gender=Male, Cy=Tape, Educaon=College Behavor=Auo Servce & Gas Saon 2 Gender=Male, Cy=Tachung, Educaon=Hgh School & Ohers Behavor= Auo Servce & Gas Saon 3 Gender=Male, Cy=Kaoshung, Educaon=Hgh School & Ohers Behavor= Auo Servce & Gas Saon Sgnfcan unexpeced changes The second ype of usage behav or pae rns s he unexpeced change. Three ypcal sgnfcan unexpeced changes are llusraed n Table 3. In he able, we found ha boh Change 1 and Change 2 are unexpeced consequen changes due o = 1. Tha s, boh rules are he same n LHS bu dfferen n RHS. For Change 1, we dscovered ha sngle women wh hgh school educaon level ncrease her spendng from 0-70,000 n year 2001 o 70, ,000 n year For Change 2, women who work n bankng wh age have her cred card usage from eneranmen n year 2001 o deparmen sore & duy free shop n year When we furher sudy he Change 2 case, s found ha more han 1/5 of cusomers n daabase (2819/10940) conform o he LHS of rules. Therefore, s srongly suggesed ha markeng persons should be aware of hs sgnfcan unexpeced change. Chang e Table 3: Sgnfcan unexpeced changes Year 2001 Year Educaon=Hgh School Educaon=Hgh & Ohers, Marrage=No, School & Ohers, Gender=Female Marrage=No, Yearly Spendng Gende r=female Amoun=0-70,000 Sup Yearly Spendng Amoun = 70, ,000 Sup Occupaon=Bankng, Occupaon=Bankng, Age= 31-40, Age= 31-40, Gender=Female Gender=Female Behavor=Eneranmen Sup Behavor= Deparmen Sore & Duy Free ShopSup Gender=Male, Occupaon=Own Gender=Male, Occupaon=Own Busness, Busness, Cy=Kaoshung Cy=Tape Behavor=Eneranmen Sup Behavor=Eneranme n Sup Sgnfcan Added/Pershed Rules The hrd ype of usage behavor change s added rules or pershed rules. Three ypcal sgnfcan added rules and hree sgnfcan pershed rules are llusraed n Table 4 and Table 5. We found ha he suppors for hose added or pershed rules are relavely low. The Rule 1 n Table 4 revealed ha cusomers lvng n Hsn-Ju wh college educaon consume on home decoraon and mprovemen n Year Alernavely, he Rule 1 n Table 5 s a sgnfcan pershed rule ndcang ha cusomers who have own-busness and lves n Pnung cy consume n deparmen sore and duy free shop n Year I s sugges ha markeng persons should noe he reasons why hose rules are added n year 2002 or no longer vald n year Table 4: Sgnfcan added rules. Rule Year 2002Added Rule Sup 1 Educaon=College, Cy=Hsnchu Behavor=Home Decoraon & Improvemen 2 Occupaon=Servce Indusry, Age= Behavor= Resauran & Pub 3 Cy=Tape, Cred Lm=200, ,000, Gender=Female Behavor=Booksore & Muscal Sore Table 5: Sgnfcan pershed rules. Rule Year 2001Pershed Rule Sup 1 Occupaon=Own Busness, Cy=Pngung Behavor= Deparmen Sore & Duy Free Shop 2 Marrage=Yes, Cy=Changhua Behav or= Deparmen Sore & Duy Free Shop 3 Cy=Tape, Age=41-50, Cred Lm= Behavor=Home Decoraon & Improvemen

6 Proceedngs of he 5h WSEAS In. Conf. on COMPUTATIONAL INTELLIGENCE, MAN-MACHINE SYSTEMS AND CYBERNETICS, Vence, Ialy, November 20-22, Conclusons In recen year, cred cards have been one of he mos popular fnancal producs. The magnfcen ncrease n cred card markes leads card ssuers pu more effors o undersand her usage behavor. To fulfll hs need, hs research proposes a deecon model o denfy usage behavor change paerns of cred card cusomers n wo me perods. Frs, cusomer profles and ransacon daa are rereved, preprocessed, and sored as wo daases. Then, assocaon rule ses are generaed for each daase usng Apror algorhm. The rules n he wo me perods are classfed as emergng paern, unexpeced change, or added/pershed rule usng defned smlary and dfference measures. Wh he proposed model, card ssuers can deec crcal usage behavor changes and allocae her lmed resource o esablsh a more suable markeng sraegy for her cusomers. The proposed model has been successfully mplemened usng real cred card daa provded by a commercal bank n Tawan. Several markeng sraeges have been suggesed n hs paper accordng o he analyss fndng. However, here are sll some rooms for model mprovemen n he fuure. Currenly, dscrezaon process n he sage wo s conduced manually by markeng persons of he card ssuer. The decson for seng appropraed numercal nerval could be very subecve. I s suggesed ha auomac dscrezaon algorhms can be appled o mprove he effcency of he model. Besdes, s worhwhle o explore varan cusomer groups and sudy how dfferen markeng sraeges wll affec her behavor. References: [1] The Nlson repor, U.S. cred and deb cards proeced hru 2007, Ocober 2003, pp [2] J. Peppard, Cusomer Relaonshp Managemen (CRM) n fnancal Servces, European Managemen Journal, Vol. 18, No. 3, 2000, pp [3] D. Peppers, M. Rogers, and R. Dorf, Is your company ready for one-o-one markeng, Harvard Busness Revew, Jan-Feb 1999, pp [4] P. Gudc, and G. Passerone, Daa mnng of assocaon srucures o model consumer behavor, Compuaonal Sascs and Daa Analyss, Vol. 38, 2002, pp [5] J. Han and M. Kamber, Daa mnng: conceps and echnques, San Francsco, Morgan Kaufmann Publshers, [6] G. Nakhaezadeh, C. Taylor, and C. Lanqullon, Evaluang usefulness of dynamc classfcaon, Proceedngs of he Fourh Inernaonal Conference on Knowledge Dscovery and Daa Mnng, [7] G. Dong, and J. L, Effcen mnng of emergng paerns: dscoverng rends and dfferences, Proceedngs of he Ffh Inernaonal Conference on Knowledge Dscovery and Daa Mnng, 1999, pp [8] B. Lu and W. Hsu, Pos-analyss of learned rules, Proceedngs of he Threen Naonal Conference on Arfcal Inellgence, 1996, pp [9] B. Lu, W. Hsu, Y. Ma, and S. Chen, Mnng neresng knowledge usng DM-II, Proceedngs of he Ffh Ineranonal Conference on Knowledge Dscovery and Daa Mnng, 1999, pp [10] J. Han, G. Dong, and Y. Yn, Effcen mnng of paral perodc paerns n me seres daabase, Proceedngs of he Ffeenh Inernaonal Conference on Daa Engneerng, 1999, pp [11] R. Agrawal and R.Srkan, Fas algorhm for mnng assocaon rules n large daabases, Proceedng of In l Conf. VLDB, 1994, pp [12] C. Lanqullon, Informaon flerng n changng domans, Proceedngs of he Inernaonal Jon Conference on Arfcal Inellgence, 1999, pp [13] H. S. Song, J. K. Km, and S. H. Km, Mnng he change of cusomer behavor n an nerne shoppng mall, Exper Sysems wh Applcaons, Vol. 21, 2001, pp