Creating Adaptive Web Sites Through Usage-Based Clustering of URLs

Transcription

1 Creating Adaptive Web Sites Through Usage-Based Clustering of URLs Bamshad Mobasher Dept. of Computer Siene, DePaul University, Chiago, IL Robert Cooley, Jaideep Srivastava Dept. of Computer Siene, University of Minnesota, Minneapolis, MN 1 Introdution Web personalization has quikly moved from an added value feature to a neessity, partiularly for large information servies and sites that generate revenue by selling produts. Web personalization an be viewed as using user preferenes profiles to dynamially serve ustomized ontent to partiular users. User preferenes may be obtained expliitly, or by passive observation of users over time as they interat with the system. Prinipal elements of Web personalization inlude modeling of Web objets (pages, et.) and subjets (users), mathing between and aross objets and/or subjets, and determination of the set of ations to be reommended for personalization. Existing approahes used by many Web-based ompanies, as well as approahes based on ollaborative filtering (e.g., GroupLens [HKBR99] and Firefly [SM95]), rely heavily on human input for determining the personalization ations. This type of input is often a subjetive desription of the users by the users themselves, and thus prone to biases. Furthermore, the profile is stati, and its performane degrades over time as the profile ages. Reently, a number of approahes have been developed dealing with speifi aspets of Web usage mining for the purpose of automatially disovering user profiles. For example, Perkowitz and Etzioni [PE98] proposed the idea of optimizing the struture of Web sites based o-ourrene patterns of pages within usage data for the site. Shehter et al [SKS98] have developed tehniques for using path profiles of users to predit future HTTP requests, whih an be used for network and proxy ahing. Spiliopoulou et al [SF99], Cooley et al [CMS99], and Buhner and Mulvenna [BM99] have applied data mining tehniques to extrat usage patterns from Web logs, for the purpose of deriving marketing intelligene. Shahabi et al [SZA97], Yan et al [YJGD96], and Nasraoui et al [NFJK99] have proposed lustering of user sessions to predit future user behavior. In this paper we desribe an approah to usage-based Web personalization taking into aount both the offline tasks related to the mining of usage data, and the online proess of automati Web page ustomization based on the mined knowledge. Speifially, we propose an effetive tehnique for apturing ommon user profiles based on assoiation-rule disovery and usage-based lustering. We also propose tehniques for ombining this knowledge with the urrent status of an ongoing Web ativity to perform realtime personalization. Finally, we provide an experimental evaluation of the proposed tehniques using real Web usage data. 2 Mining Usage Data for Web Personalization The overall proess of usage-based Web personalization an be divided into two omponents. The offline omponent is omprised of the data preparation tasks resulting in a user transation file, and the speifi usage mining tasks, whih in our ase inlude the disovery of assoiation rules and the derivation of URL lusters based on user aess patterns. One the mining tasks are aomplished, the frequent itemsets and the URL lusters are used by the online omponent of the arhiteture to provide dynami reommendations to users based on their urrent navigational ativity. The Web server keeps trak of the ative user session as the user browser makes HTTP requests. The reommendation engine onsiders the ative user session in onjuntion with the URL lusters to ompute a set of reommended URLs. The reommendation set is then added to the last requested page as a set of links before the page is sent to the lient browser. A generalized arhiteture for 1

2 Figure 1. General Arhiteture for Usage-Based Web Personalization usage-based personalization is depited in Figure 1. In this setion, we disuss the offline omponents of this arhiteture. The online ustomization proess is presented in the next setion. 2.1 Data Preparation Tasks A ritial step in effetively mining usage data for Web personalization is the leaning and transformation of aess log data, and the identifiation of a set of user sessions. Cleaning the server logs involves removing redundant referenes (e.g., image and sound files, multiple frames, and dynami pages that have the same template), leaving only one entry per page view. It is also neessary to filter the log files by mapping the referenes to the site topology indued by physial links between pages. Client-side and proxy level ahing often reate impediments to the identifiation of unique user sessions. For example, in a Web server log, all requests from a proxy server have the same identifier, even though the requests potentially represent more than one user. Tehniques suh as the use of lient-side ookies for user identifiation are not always pratial due to privay onerns of the users, or limitations of the apabilities of the Web server. In [CMS99] we proposed several simple heuristis using the referrer and agent fields of a Server log to identify user sessions and infer missing referenes with relative auray in the absene of additional information suh as ookies. The goal of transation identifiation is to dynamially reate meaningful lusters of referenes for eah user. Based on an underlying model of the user's browsing behavior, eah page referene an be ategorized as a ontent referene, auxiliary (or navigational) referene, or hybrid. In this way different types of transations an be obtained from the user session file, inluding ontent-only transations involving referenes to ontent pages, and navigation-ontent transations involving a mix of pages types. The details of 2

3 methods for transation identifiation are disussed in [CMS99]. Finally, the session file may be filtered to remove small transations and low support referenes (i.e., URL referenes that are not supported by a speified number of user transations). This type of support filtering is important in removing noise from the data, suh as transations orresponding to users who do not traverse the site. Given the preproessing steps outlined above, for the rest of this paper we assume that there is a set of n unique URLs U = {url 1, url 2,, url n }, appearing in the preproessed log, and a set of m user transations T = {t 1, t 2,, t m }, where eah t i T is a non-empty subset of U. 2.2 Usage Clustering Based on Frequent Itemsets Traditional ollaborative filtering tehniques are often based on mathing the urrent user's profile against lusters of similar profiles obtained by the system over time. Similar tehniques have been proposed for obtaining user profiles from Web usage data by lustering user sessions [SZAS97, YJGD96]. However, session lusters by themselves are not an effetive means of apturing an aggregated view of ommon user aess patterns. Eah session luster may potentially ontain thousands of user sessions involving hundreds of URL referenes. Furthermore, the omputation of similarity or distane between user session vetors in the ontext of Web usage mining is not a trivial task, sine feature weights based on the amount of time spent by a partiular user on a page, or the number of referenes to a page within a session, are generally not good behavioral indiators in Web transations. In ontrast to lustering user sessions, we diretly ompute overlapping lusters of URL referenes based on their o-ourrene patterns aross user transations. We all the URL lusters obtained in this way, usage lusters. Our experiments suggest that usage lusters tend to group together related items (based on user aess patterns) aross transations, even if these transations are themselves not deemed to be similar. This allows us to obtain lusters that potentially apture overlapping interests of different types of users. However, traditional lustering tehniques, suh as distane-based methods generally annot handle this type lustering. The reason is that instead of using URLs as features, the transations must be used as features, whose number is in tens to hundreds of thousands in a typial appliation. Furthermore, dimensionality redution in this ontext may not be appropriate, as removing a signifiant number of transations as features will result in losing too muh information. Assoiation Rule Hypergraph Partitioning (ARHP) tehnique [HKKM97, HKKM98] is well-suited for this task, sine it provides automati filtering apabilities, does not require distane omputations, and an be used in high-dimensional data sets without requiring dimensionality redution. The ARHP tehnique has been used suessfully in a variety of domains, suh as ontent-based ategorization of Web douments [HBG+99]. Assoiation rules apture the relationships among items based on their patterns of o-ourrene aross transations. In Web transations, assoiation rules apture relationships among URL referenes based on the navigational patterns of users. The assoiation rule disovery methods, suh as the Apriori algorithm [AS94], initially find groups of items (whih in this ase are the URLs appearing in the preproessed log) ourring frequently together in many transations. Suh groups of items are referred to as frequent item sets. Given a set I = {I 1, I 2,, I k } of frequent itemsets, the support of I i is the fration of transations ontaining I i, and is denoted by σ(i i ). Generally, a support threshold is speified before mining and is used by the algorithm for pruning the searh spae. An assoiation rule r is an expression of the form <X Y, σ r, α r >, where X and Y are sets of items, σ r is the support of X Y, and α r is the onfidene for the rule r given by σ(x Y) / σ(x). The frequent itemsets are used as hyperedges to form a hypergraph H = <V, E>, where V U and E I. A hypergraph is an extension of a graph where eah hyperedge an onnet more than two verties. The weights for hyperedges are omputed based on the onfidene of the assoiation rules involving the items in the frequent itemset. The hypergraph H is then partitioned into a set of lusters C = { 1, 2,, h }. The similarity among items is aptured impliitly by the assoiation rules. Eah luster is examined to filter out verties that are not highly onneted to the rest of the verties of the partition. The onnetivity of vertex v (a URL appearing in the frequent itemset) with respet to a luster is defined as follows: onn( v, ) = { e e, v e} { e e } 3

4 Connetivity measures the perentage of edges with whih a vertex is assoiated. A high onnetivity value suggests that the vertex has many edges, onneting a good proportion of the verties in the partition. The verties with onnetivity measure greater than a given threshold value are onsidered to belong to the luster. Additional filtering of non-relevant items an be ahieved using the support riteria in the assoiation rule disovery omponents of the algorithm. One the URL lusters have been omputed, a partial session for the urrent user an be assigned to a mathing luster. The onnetivity value of an item (URL) defined above is important beause it is used as the weight assoiated with that item for the luster. These weights are used as part of the reommendation proess when lusters are mathed against an ative user session. The details of the mathing and reommendation proess are disussed in the next setion. 3 Automati Customization Based on Usage Clusters The online omponent of the Web personalization system involves the omputation of a reommendation set for the urrent session, onsisting of links to pages that the user may want to visit based on similar usage patterns. The reommendation set essentially represents a "short-term" view of potentially useful links based on the user's navigational ativity through the site. These reommended links are then added to the last page in the session aessed by the user before that page is sent to the browser. We use a fixed-size sliding window over the ative session to apture the urrent user s history depth. A sliding window of size n over the ative session allows only the last n visited pages to influene the omputation of the reommendation set. This makes sense in the ontext of personalization sine most users go bak and forth while navigating a site to find independent piees of information. In many ases these subsessions have a length of no more than 2 or 3 referenes. The notion of a sliding session window is similar to the notion of N-grammars disussed in [Cha96]. We an determine the optimum window size based on the average user transation length identified during the preproessing stage. We also want to weight a URL reommendation higher, if it is farther away form the urrent ative session. To apture this notion, we maintain a direted graph, G, representing the topology of the site. The physial link distane between two URLs u 1 and u 2 is the length of a minimal path from u 1 to u 2 in this site graph. Now, the physial link distane between the ative session s and a URL u s is denoted by dist(u,s,g), whih is defined as the smallest physial link distane between u and any of the URLs in s. The link distane fator of u with respet to s is defined as ldf(u,s) = log(dist(u,s,g))+1. If the URL u is in the ative session, then ldf(u,s) is taken to be 0. We take the log of the link distane so that it does not ount too heavily ompared to item weights within lusters. Eah URL luster an be viewed as virtual user profile representing ommon aess patterns. One a new user starts a session, our goal is to math, at eah step, the partial user session with the usage lusters, and provide dynami reommendations to the user. Reall that the weight of URL u in a luster is the onnetivity value of the item within the luster, denoted by onn(u, ), as defined in the previous setion. We an therefore represent eah luster C, as a vetor G u, u,!, u, where > = 1 2 onn( urli, ), if urli u i = 0, otherwise The urrent ative session s is also represented as a (binary) vetor <s 1, s 2,, s n >, where s i = 1, if the user has aessed URL i in this session, and s i = 0, otherwise. The luster mathing sore is now omputed as follows: uk sk k math( s, ) = s ( u ) 2 k k Note that the mathing sore is normalized for the size of the lusters and the ative session. This orresponds to the intuitive notion that we should see more of the user's ative session before obtaining a better math with a larger luster representing a user profile. Given a luster and an ative session s, a reommendation sore, Re(s,u), is omputed for eah URL u in as follows: n 4

5 Re( s, u) = onn( u, ) math( s, ) ldf 0s, u5. If the URL u is in the urrent ative session, then its reommendation value is zero beause of the link distane fator. Finally, we an ompute the reommendation set, Reommend(s), for urrent ative session s by olleting from eah luster all URLs whose reommendation sore satisfies a minimum reommendation threshold ρ, i.e., Reommend ( s) = { u C, and Re( s, u ) ρ }. For eah URL that is ontributed i i by several lusters, we use its maximal reommendation sore from all of the ontributing lusters. 4 Experimental Results We used the aess logs from the University of Minnesota Computer Siene Web server to evaluate our personalization tehnique. The preproessed log (for February of 1999) was onverted into a session file omprising user transations and a total of 4001 unique URLs (before support filtering). We used the hypergraph partitioning algorithm as modified in [CC99] in order to take frequent itemsets as the input, and performed the lustering of URLs. The frequent itemsets were found using the tree projetion algorithm desribed in [AAP99]. Eah URL serves as a vertex in the hypergraph, and eah edge represents a frequent itemset with the weight of the edge taken as the interest for the set. Sine interest inreases dramatially with the number of items in a rule, the log of the interest is taken in order to prevent the larger rules from ompletely dominating the lustering proess. For the reommendation proess we hose a session window size of 2, sine the average session size was 2.4. The reommendation results are given for the sample path /researh=>/grad-info=>/registration-info. A summary of the results are given in Figure 2; additional experimental results and a demonstration site using the tehniques disussed in this paper an be found at Eah table in Figure 2 orresponds to one step in the user navigation through the path. In eah ase the urrent ative session window is given along with the top reommendations. A ut-off value of 0.30 was used for the reommendation sore. Eah row shows two onseutive reommendations. Note that in many ases the obvious reommendations assoiated with the URLs in the ative session window rank lower than URLs for pages that are farther away in the site graph. This variation is mainly due to the link distane fator disussed earlier. In eah ase the reommendation set is omposed of URLs from a number of mathing lusters. When /researh page is requested, the URLs for a number of popular researh groups are inluded. But, note that pages from graduate information and registrations setions are also inluded, sine many users (mainly graduate students) often have overlapping interests aross these setions. When /grad-info is requested, some of the frequently visited URLs assoiated with that page as well as related lass registration pages rank higher. Finally, when /registration-info is requested, a more foused set of reommendations results, not involving researh related pages. 5 Conlusions The ability to ollet detailed usage data at the level of individual mouse liks, provides Web-based ompanies with a tremendous opportunity for personalizing the Web experiene of lients. In e-ommere parlane this is being termed mass ustomization. Most urrent approahes to personalization by various Webbased ompanies rely heavily on human partiipation to ollet profile information about users. This suffers from the problems of the profile data being subjetive, as well getting out of date as the user preferenes hange over time. In this paper we have presented an arhiteture for automati Web personalization based on Web usage data. We introdued an effetive lustering tehnique using assoiation rule mining to learn overlapping user profiles, and disussed how the extrated knowledge an be used in real-time to provide navigational pointers for users. Our experimental results indiate that the tehniques disussed here are promising, and bear further investigation and development. Our future work in this area involves onduting experiments with various types of transations derived from user sessions, for example, to isolate speifi types of "ontent" pages in the reommendation proess. We also plan on inorporating lient-side agents to provide an additional level of personalization based on user preferenes. 5

6 Session Window: /researh Reommendation Sore Reommendation Sore /newsletter/newfaulty.html 0.73 /newsletter 0.65 /faulty 0.55 /researh/nmrg 0.55 /researh/softeng 0.55 /researh/airvl 0.51 /researh/mmdbms 0.48 /researh/agassiz 0.47 /personal-pages 0.37 /registration-info 0.35 /registration-info/spring99.html 0.32 /grad-info 0.30 /registration-info/sh99-00.html 0.30 /grad-researh 0.30 Session Window: /researh => /grad-info Reommendation Sore Reommendation Sore /faulty 0.59 /personal-pages 0.52 /newsletter/newfa.html 0.52 /newsletter 0.46 /grad-info/grad-handbook.html 0.45 /grad-info/ourse-guide.html 0.45 /grad-info/prospetive-grads.html 0.40 /registration-info 0.39 /researh/nmrg 0.39 /researh/softeng 0.39 /researh/airvl 0.36 /registration-info/spring99.html 0.35 /researh/mmdbms 0.34 /researh/agassiz 0.33 Session Window: /grad-info => /registration-info Reommendation Sore Reommendation Sore /faulty 0.51 /personal-pages 0.45 /grad-info/grad-handbook.html 0.45 /grad-info/ourse-guide.html 0.45 /grad-info/prospetive-grads.html 0.40 /registration-info/spring99.html 0.36 /registration-info/sh99-00.html 0.34 Figure 2. Reommendation Results for a Typial User Navigation Path Referenes [AAP99] [AS94] [BM99] Agarwal, R., Aggarwal, C., and Prasad, V., A tree projetion algorithm for generation of frequent itemsets. In Proeedings of High Performane Data Mining Workshop, Puerto Rio, Agrawal, R. and Srikant, R., Fast algorithms for mining assoiation rules. In Proeedings of the 20 th VLDB onferene, pp , Santiago, Chile, Buhner, A. and Mulvenna, M. D., Disovering internet marketing intelligene through online analytial Web usage mining. SIGMOD Reord, (4) 27, [Cha96] Charniak, E., Statistial language learning. MIT Press, [CC99] [CMS99] [CPY96] [HBG+99] [HKBR99] Clifton, C. and Cooley, R., TopCat: data mining for topi identifiation in a text orpus. In Proeedings of the 3rd European Conferene of Priniples and Pratie of Knowledge Disovery in Databases, Prague, Czeh Republi, Cooley, R., Mobasher, B., and Srivastava, J., Data preparation for mining World Wide Web browsing patterns. Journal of Knowledge and Information Systems, (1) 1, Chen, M. S., Park, J. S., and Yu, P. S., Data mining for path traversal patterns in a Web environment. In Proeedings of 16th International Conferene on Distributed Computing Systems, Han, E-H, Boley, D., Gini, M., Gross, R., Hastings, K., Karypis, G., Kumar, V., and Mobasher, B., More, J., Doument ategorization and query generation on the World Wide Web using WebACE. Journal of Artifiial Intelligene Review, January Herloker, J., Konstan, J., Borhers, A., Riedl, J.. An algorithmi framework for performing ollaborative filtering. To appear in Proeedings of the 1999 Conferene on Researh and Development in Information Retrieval, August

7 [HKKM97] [HKKM98] [NFJK99] [PE98] [SF99] [SKS98] [SM95] [SZAS97] [YJGD96] Han, E-H, Karypis, G., Kumar, V., and Mobasher, B., Clustering based on assoiation rule hypergraphs. In Proedings of SIGMOD 97 Workshop on Researh Issues in Data Mining and Knowledge Disovery (DMKD 97), May Han, E-H, Karypis, G., Kumar, V., and Mobasher, B., Hypergraph based lustering in highdimensional data sets: a summary of results. IEEE Bulletin of the Tehnial Committee on Data Engineering, (21) 1, Marh Nasraoui, O., Frigui, H., Joshi, A., Krishnapuram, R., Mining Web aess logs using relational ompetitive fuzzy lustering. To appear in the Proeedings of the Eight International Fuzzy Systems Assoiation World Congress, August Perkowitz, M. and Etzioni, O., Adaptive Web sites: automatilly synthesizing Web pages. In Proeedings of Fifteenth National Conferene on Artifiial Intelligene, Madison, WI, Spiliopoulou, M. and Faulstih, L. C., WUM: A Web Utilization Miner. In Proeedings of EDBT Workshop WebDB98, Valenia, Spain, LNCS 1590, Springer Verlag, Shehter, S., Krishnan, M., and Smith, M. D., Using path profiles to predit HTTP requests. In Proeedings of 7th International World Wide Web Conferene, Brisbane, Australia, Shardanand, U., Maes, P., Soial information filtering: algorithms for automating "word of mouth." In Proeedings of the ACM CHI Conferene, Shahabi, C., Zarkesh, A. M., Adibi, J., and Shah, V., Knowledge disovery from users Web-page navigation. In Proeedings of Workshop on Researh Issues in Data Engineering, Birmingham, England, Yan, T., Jaobsen, M., Garia-Molina, H., Dayal, U., From user aess patterns to dynami hypertext linking. In Proeedings of the 5 th International WWW Conferene, Paris,