Web Usage Patterns Using Association Rules and Markov Chains

Transcription

1 Web Usage Paerns Using Associaion Rules and Markov hains handrakasem Rajabha Universiy, Thailand Absrac - The objecive of his research is o illusrae he probabiliy of web page using a a period of ime using wo saisical echniques. SME Nonhaburi province handicraf goods e-commerce web sie was seleced as a case sudy in his research. Web pages were caegorized ino hree porions: SME firm secion s News, Goods deails and usomer aciviies. Markov chain echnique was applied in order o presen probabiliy of each even. Associaion rule echnique was also used o derive he pahs of web page visiing. This complemenary resuls from wo echniques should suppor web adminisraor o spo wha web pages are he ineresed web pages of his web sie archiecure. Associaion rules show he pah of consequence visied web page while Markov chains give and informaion of possibiliy of evens ha should be visied in specific assigned period of ime. Keywords - Web Mining, Markov hains, Associaion Rule I. INTRODUTION A websie adminisraor (also known as admin) has responsibiliies o creae, mainain and fulfill he imporan daa and informaion of organizaion in order o suppor he enerprise mission. Some conens are ineresed bu anoher one are bored. The conens ha are mosly simulaneous visiing may cause he downgrade of sysem performance. Adminisraor mus deeply analyze he sofware design, daabase design, server or even nework bandwidh adjusmen in order o address hese problems. In case of less visied web page, adminisraor has o modify, increase some conens or deleed hem. This research gahered web usage from sys-log daabase. The e-commerce websie was used as an experimen. The conen in websie was caegorized ino hree ypes such as SME firm secion s News (even A), Goods deails (even B) and usomer aciviies (even ). The 12 web usage daa observaions were gahered during 1-31January 218. These observaions were used o find ou heir paerns of web page visiing by using Associaion rule echnique. Meanwhile, Markov chains were also calculaed in order o obain he probabiliy of evens in defined period of ime. From he resuls, boh echniques provide informaion abou web usage paerns and possibiliy of occurring so ha web sie adminisraor can use hese in conen managemen. II. RELATED THEORY AND RESEARH A. Relaed Research 1) Associaion rule echnique was applied in web usage mining. Observaions were gahered from web usage log file of web page VTSNS, he Advanced School of Technology Novi Sad Serbia, web sie. The experimen was repeiive pruning huge received rules by seing value of suppor and confidence. The derived rules were used in web sie map (archiecure) modificaion. 2) Two web browsers, under exposed, ha were used and experimened in he server of Deparmen of Mahemaics and Saisics, Sagar Universiy ha which one was mos populariy and appreciae. Markov chains model was used o illusrae probabiliy of heir wo browsers usage afer passing ime uni 47

2 Web Usage Paerns Using Associaion Rules and Markov hains afer sar ime. The less preferred browser, small value probabiliy of exisence, was inspeced for is problems, performance, ec. Afer his less popular web browser was modified abou undesired feaures, such as adjus-add on-re configuraion, Markov chains was hen re processing. The modified web browser was become more increase in probabiliy of sae exisence han before. B. Associaion Rule [3] Associaion rule is a saisical echnique ha used o find ou he dependency of aribues. For example, if aribue A is occurred while aribue B is also occurred hen i can define he rule as A B. The crieria of decision making from he rule can be acceped are wo merics: suppor and confidence ha can be compued by (1) and (2). sup por( A B) ( A B) (1) n( A B) confidence ( A B) (2) n( A) where n( A B) is he number of observaions ha A (source) and B (desinaion) are boh occurred. na ( ) is oal observaions ha sae A is presence. Normally, a suppor value is calculaed from oal daa se herefore his value is used o find ou he pahs have he frequenly happening of user s web sie raveling. If he suppor value is oo high hen here may a few rules are discovered. On he oher hand, here may ge many rules if he suppor value is oo small. For more deail, he confidence value covers he oal observaions ha aribue A, or source B aribues, is occurred. Therefore, he confidence value presens he probabiliy of specific rules ha he happening of desinaion aribues when he source aribues of observaions are oally occurred. If his value is high, i means ha hese rules have more happening in case of he specific occurring source aribue observaions daase.. Markov hains [4] Markov chains is a saisical echnique ha is used o calculae for he probabiliy of ransiion of wo evens beween wo periods in a ime. There should have more saes in he sudying problems. Thus, he specific sae could ravel o anoher sae, or even iself, under prior probabiliy of ransiion marix (P). A any ime period passed from sar poin, he probabiliy of all saes could be calculaed from he sysem of firs order difference equaion. Le is a n 1 size of vecor ha describes he possibiliy of all saes being a ime. describes he possibiliy of all saes being a he ime. The should be calculaed by (3). P 1 (3) is an iniial possibiliy vecor of all saes. I is used o derive he value of all parameers in a sysem of he firs order difference equaion ha is used o find ou all possible evens a a period of ime wihou equaion (3) in consequence from =, 1, 2, 3, -1. Thus, he firs order difference equaion provides more comforable in compuing han ypical processing. III. RESEARH METHODOLOGY A. Daa Preparaion Daa Source: A number of 12 records of he observaion were colleced from SME Nonhaburi Province handicraf goods e-commerce Websie. Websie map is composed of hree menus. The firs porion explains abou he mission of privae enerprise. The second porion is an imporan menu since i gives informaion abou enerprise s goods. And he hird porion informaion and aciviies of order, paymen and goods receive. Each menu is composed of many sub menus hus his research considered only in hree groups in 48

3 order o reduce compuaional complexiy. Web Usage Log: Our research case sudied web sie applicaion has designed a daabase ha was used o keep all users acions a he choosing menu from sar even, or sae (invocae), and oher raveling saes unil hey logoff he websie. TABLE I DATA DETAIL OF WEB USAGE LOG DATABASE Aribues Descripion Daa Type IP address Dae Time Menu-chosen IP address of exernal user Day: monh: year Hour: min Menu ype S=sar, L=logoff, A=Firm s news, B=Goods deail, =usomer aciviy :99:99 99:99 S, A, B,, L Noe: Daa collecion period: Web usage logs were gahered during 1-3 January 218. B. Associaion Rules From web usage log daa base, each user s daa, observaion, were coding and cleaning before furher daa processing sep. Objecive of daa preparaion was o presen he absence and presence of even, or sae, during period of ime, for an example, five observaions were shown in Table II. If he user selecs any menu (S, A, B,, L) hen chosen menu wascoded as 1 (presence), The was done if he absence. TABLE II PARTIAL DATA ABOUT USER S MENU SELETION Observaion# Sar A B Logoff Some combinaion of pahs do no happen such as L A, A S Thus his kind of pah mus be deleed. The amoun of all possible pahs, or rule, in an Associaion rule is shown in (4). d d1 # ofpossiblerule (4) While d is an amoun of even, or aribue, an ineres experimen. For example, if here are 3 evens as: A - B -, hen here are 12 possible rules. A B, A, A B, B, B A, B, A, B A, B, A, B, B, A In his research, here were five evens (d=5) hen here were one hundred possible associaion rules. Some of hese rules were measured by suppor and confidence merics while some pahs did no pass. In order o address he problem of numerous rules abou imporan and less imporan conens. Thus, he suppor and confidence values should be he high score. In his research, he accepable suppor value was se o.3 and he confidence value was se o.5. There were four discovered rules ha passed boh crieria (descending order) as shown in Table III. The mosly happen pah is S A L, suppor =.33. Many cusomers sar visi his websie, read News hen log off. The enerprise goods migh no be ineresed so ha hey suddenly leave his web sie. TABLE III USER S MOSTLY OURRED PATHS # Suppor onfidence Rule or Pah S A L S B L S B L. Markov hain All daa of user s menu selecion, during hey spen heir ime in he websie, were summarized in working able, as shown in Table IV). For example - according o Table 49

4 Web Usage Paerns Using Associaion Rules and Markov hains IV, afer login o he websie, he 3% of users (p=.3) chose o visi menu A. And he logoff was chosen a 1%. In case of A sae, afer A was chosen, 1% of user sill chose o say in A and he 3% chose o ravel o B,, and L (logoff). The experimenal ransiion probabiliies marix ( P ) deail is shown in Table IV. TABLE IV SUMMARY PROBABILITY OF USERS BEHAVIORS Sar A B Logoff Sar A B Logoff According o ransiion probabiliy marix, here were connecing pah beween some node, or sae, wih oher nodes or even is self while some connecing pahs were absence since here were no ransiion probabiliy beween hem. The whole relaion pahs were illusraed, as shown in Fig. 1. S.2.1 are cumbersome in calculaion cause he nex probabiliy of ineresed even is depend on prior even probabiliy hus hese can made simplified by anoher echnique, he sysem of firs order difference equaion mehod. Afer he daa preparaion was finished, i was summarized ha iniial sae ( ) has probabiliy column vecor as shown in (5), he probabiliy even vecor a ime as shown in (6) and he experimenal ransiion probabiliy marix ( P ) as shown in (7) S A B L (5) (6) P (7).1 A.4 B.2.2 Afer a long rial of mahemaical calculaion, he sysem of firs order difference equaion for all evens a ime period was compued hen all even s probabiliy or vecor was presened in equaion (8), (9), (1), (11), and (12). L S.25.16(.56).3(.2).7(.2).2(.21) (8) Fig. 1 Markov Model of Transiion Probabiliy Marix (P) From Table IV, daa was called ransiion probabiliy Marix ( P ). This marix presens an explanaion abou users behaviors abou he web sie s menu choosing. Markov chains mehod was hen applied o illusrae he model of all probably evens (or sae) in a specific ime period. Markov chains A.12.15(.56).15(.2).1(.2).11(.21) (9) B.21.33(.56).3(.2).3(.2).34(.21) (1).9.12(.56).12(.2).4(.2).18(.21) (11) L.48.44(.56).3(.2).1(.2).2(.21) (12) 5

5 IV. RESEARH SUMMARY AND SUGGESTION A. Summary The proposed research echniques in his research could presen he co-occurrence among evens under he arbirary defined dependency level such as suppor and confidence. Websie adminisraor can choose he value of rules as well as ha wheher resul rules should be sufficien o explain he cusomers behaviors or no. According o he calculaion resuls, Markov chains can presen he probabiliy of paricular evens. While he associaion rules give all possible consequence pahs ha are relaed o he occurred evens bu no any occurrence of probabiliy abou each even. Significance associaion rule pah informs web sie owner abou user or cusomers behaviors. Pah #2 and #3 presen how ofen users visi websie from menu A, B hen go o L. Some cusomers visi he evens B and hen go o L. In pracice, if here is an amoun of oal cusomers N ha coun he number of websie visiing a a period of ime, such as he firs day of any monh, hen he roughly possible amoun of cusomers M ha should follow he pah #3 ha could prediced by (13). funcion, Informaion heory, ausal model analysis and ec should be considered as oher possible echniques ha can be used o solve his problem. REFERENES (Arranged in he order of ciaion in he same fashion as he case of Foonoes.) [1] Dimirijević, M. (211). Web Usage Associaion Rule Mining Sysem. Inerdisciplinary Journal of Informaion, Knowledge, and Managemen, Vol. 6. [2] Shukla, D. (211). Analysis of Users Web Browsing Behavior Using Markov chain Model. Deparmen of Mahemaics and Saisics, Sagar Universiy, Sagar M.P., 473, India. [3] Kumar, V. (25). Inroducion o Daa Mining. Pang-Ning Tan, Michael Seinbach, Vipin Kumar Addison-Wesley, ISBN: [4] Fewser. Markov chains. Auckland Universiy, New Zealand. M N * B * (13) The main objecive of e-commerce web sie is o provide he ineresing goods o huge number of cusomers so ha many poenial visiors decide o buy web sie s goods. Therefore, he percenage of web sie success could be calculaed from (14). M SuccessPercenage *1% (14) N B. Furher Research Based on echniques ha were used here, all significan even dependences migh (or no) mee he saisical significance crieria (.5 ) since hey give no any informaion abou ype I error. In alernaive, Bayesian heorem under join probabiliy densiy 51