ARS: Web Page Recommendation System for Anonymous Users Based On Web Usage Mining

Transcription

1 ARS: Web Page Recommendation System for Anonymous Users Based On Web Usage Mining Yahya AlMurtadha, MD. Nasir Bin Sulaiman, Norwati Mustapha, Nur Izura Udzir and Zaiton Muda University Putra Malaysia, UPM Serdang, Selangor Darul Ehsan, Malaysia Tel: , Fax: y.murtadha@gmail.com, {nasir, norwati, izura,zaiton}@fsktm.upm.edu.my Abstract Web now becomes the backbone of the information. Today the major concerns are not the availability of information but rather obtaining the right information. Mining the web aims at discovering the hidden and useful knowledge from web hyperlinks, contents or usage logs. This paper focuses on improving the prediction of the next visited web pages and recommends them to the current anonymous user based on web usage mining technique where many data mining techniques applied to web server logs. We proposed ARS to recommend to the anonymous web user by assigning him to the best navigation profiles obtained by previous navigations of similar interested users based on his early stage navigation. To represent the anonymous user s navigation history, we used a window sliding method with size n over his current navigation session. Using CTI dataset the experimental results show higher prediction accuracy for the next visited pages for anonymous users compared to previous recommendation system. Keywords Web Usage Mining, Recommendation Systems, Usage Profiling. T I. INTRODUCTION HE rapid growth of the Web in the last decade makes it the largest publicly accessible data source in the world. The web now becomes one of the main sources of the information. Now it includes more than 4 billion pages, with about 1 million added every day [1]. Web mining aims to discover useful information or knowledge from Web hyperlinks, page contents, and usage logs[2]. Yet an important problem is how to mine complex data formats including Image, Multimedia, and Web data [3]. Based on the primary kinds of data used in the mining process, Web mining tasks can be categorized into three main types: Web structure mining, Web content mining and Web usage mining[2]. Web structure mining discovers knowledge from hyperlinks, which represent the structure of the Web. Web content mining extracts useful information/knowledge from Web page contents. Web usage mining (WUM) mines user access patterns from usage logs, which record clicks made by every user. Fig.1 shows the WUM recording process of the users browsing activities either from direct client-server browsing or proxy-server browsing in terms of IP address, date, method, file required, protocol, browser types, et which stored at the web server logs files. The goal is to capture, model, and analyze the behavioral patterns and profiles of users interacting with a Web site. The discovered patterns are usually represented as collections of pages, objects, or resources that are frequently accessed by groups of users with common needs or interests. These patterns are discovered by applying some clustering algorithms on the preprocessor phase of the web usage mining and classification algorithms on the web mining process. The output of the WUM is some patterns that may be the input to the Recommendation systems Engine which is one of the application areas of the Web usage gives the ability to predict the next visited page for a given user. Fig. 1 Web Usage Mining Recording Process of the users browsing Activities Manuscript received September 30, Yahya AlMurtadha, PhD candidate, faculty of Computer Science and Information Technology, University Putra Malaysia, UPM Serdang, Selangor Darul Ehsan, Malaysia. Phone: , Fax: Associate Professor Dr. Md. Nasir Bin Sulaiman is the leader of the Intelligent Computing Research Group at the faculty of Computer Science and Information Technology, University Putra Malaysia. Norwati Mustapha, Nur Izura Udzir and Zaiton Muda are lecturers at the department of Computer Science, faculty of Computer Science and Information Technology, University Putra Malaysia II. RELATED WORKS The major application areas for Web usage mining fall into five categories: personalization, system improvement, site modification, business intelligence, and usage characterization [4]. Recently, many researches tried to improve the prediction accuracy of the recommendation systems; however the current recommendation systems can not satisfy the users especially with the increasing growth of the web sites, the increasing ISBN:

2 number of the web users and the preferences changing of these users at any time and most of them are user-centered recommendation systems [5]. Analog [6] is structured according to an off-line component to build session clusters and an online component to build active user sessions which are then classified according to the generated model. The classification allows returning the requested page with a list of suggestions related to the ones in the active session and to. The geometrical approach used for clustering is affected by several limitations, related to scalability and to the effectiveness of the results found. In [7] [8] Mobasher et al., present WebPersonalizer a system which provides dynamic recommendations, as a list of hypertext links, to users. In [9] Mobasher et al., presented and experimentally evaluate two techniques, based on clustering of user transactions and clustering of pageviews, in order to discover overlapping aggregate profiles that can be effectively used by recommender systems for real-time Web personalization. The prediction engine has to make a recommendation list to the user session from multiple profiles based on match score that must exceed the threshold. [10] proposed an intelligent web recommender system known as SWARS (Sequential Web Access based Recommender System) that uses sequential access pattern mining. Liu and Keselj [11] proposed a novel approach to classifying user navigation patterns and predicting users future requests based on the combined mining of Web server logs and the contents of the retrieved web pages. Baraglia and Palmerini [12] [13] proposed a WUM system called SUGGEST, that provide useful information to make easier the web user navigation and to optimize the web server performance. Potential limitation of this architecture might be: a) the memory required to store Web server pages is quadratic in the number of pages. [14] developed a conceptual framework for recommending one-and-only items. It uses fuzzy logic, which allows to reflect the graded/uncertain information in the domain, and to extend the CF paradigm, overcoming limitations of existing techniques. A possible application in the context of trade exhibition recommendation for e-government is discussed to illustrate the proposed conceptual framework. [5] Proposed WebPUM, a recommender system based on Common Sequences algorithm (LCS). [15] presented NEWER as a usage-based Web recommendation system that exploits the potential of Computational Intelligence techniques to dynamically suggest interesting pages to users according to their preferences. All of these works attempt to find architecture and algorithm to improve accuracy of personalized recommendation, but the accuracy still does not meet satisfaction. Also almost all of them are user-centered prediction engine which concentrate on recommending the next visited pages based on the previous navigation of the user. In our work we advance architecture to recommend base on the current navigation pattern of the user for improving accuracy of recommendation. A. CTI Dataset III. THE METHODOLOGY This data set contains the preprocessed and filtered sessionized data for the main DePaul CTI Web server ( The data is based on a random sample of users visiting this site for a 2 week period during April of The original (unfiltered) data contained a total of sessions from 5446 users. The filtered data files were produced by filtering low support pageviews, and eliminating sessions of size 1. The filtered data contains sessions and 683 pageviews. B. The Architecture Overview Following the standard data mining process [16], the overall Web usage mining process can be divided into three interdependent stages: data collection and pre-processing, pattern discovery, and pattern analysis. In the pre-processing stage, the clickstream data is cleaned and partitioned into a set of user transactions representing the activities of each user during different visits to the site. In the pattern discovery stage, statistical, database, and machine learning operations are performed to obtain hidden patterns reflecting the typical behavior of users, as well as summary statistics on Web resources, sessions, and users. In the final stage of the process, the discovered patterns and statistics are further processed, filtered, possibly resulting in aggregate user models that can be used as input to applications such as recommendation engines. As shown in Fig.2, the proposed architecture consists of two main components, namely the offline and online. In the offline component tow three important processes are taken. First, preprocess the web server logs by allying data cleaning techniques and then partition the web navigations into sessions determined by the period of browsing. Second, partition the filtered sessionized page views into clusters of use s navigation patterns with similar pageviews browsing activities. Finally, generate web navigation profiles based on the preformed clusters. The online component does the matching of the new anonymous user request (current active session) to the profile shares common interests to the user. The details will be discussed in the following sections. 1) Offline Component This section will describe the three steps required for the offline component, namely preprocessing the data, clustering and the building the navigation profiles. a) Preprocessing Preprocessing consists of data cleaning and Pageviews identification and Sessionization. ISBN:

3 Fig. 2 the Proposed Architecture Overview (1) Data Cleaning. The web user log files are raw in the original format; hence, many cleaning tasks should be implemented to eliminate the irrelevant and useless log entries for web usage mining. Cleaning aims at removing non-important objects references like crawlers, web robots, graphics, or sound files and other irrelevant files. pageview duration of the last pageview in that session, was estimated to be the average duration of that pageview across all sessions (in which the pageview does not occur as the exit page). The file contains sessions with 682 pageviews visited by different users. The output is K-clusters (i.e 10 clusters) each contains sessions with similar pageviews. This input will be the input to the next step (building the navigation profiles). (2) Pageviews identification and Sessionization. Pageview is a request to load a single page from an internet site. A sequence of Pageviews by a single user during a single visit is called a session. World Wide Web (W3C) has defined a session as the group of browsing activities performed by a single user from the moment he starts to the time he left. Sessionization is the process of segmenting the activity record of each user into sessions, each representing a single visit to the site. b) Clustering One important use of clustering in Web usage mining is aimed at finding groups which share common interests and behaviors by analyzing the data collected in Web servers. We used K-Mean clustering algorithm to cluster the preprocessed and filters web server logs with different K values. For the clustering purpose, we used CTI.std file as an output to the K- Means clustering algorithm. The file represents a sessionpageview matrix where each column is a pageview and each row is a session represented as a vector. The entries in the table correspond to the amount of time (in seconds) spent on pageviews during a given session. The pageview durations were maxed out at 999 seconds. For each session, the c) Building the navigation Profiles The discovery of patterns from usage data by clustering the web transaction into clusters of user sessions or pages, by itself is not sufficient for performing the personalization tasks [9]. The critical step is the effective derivation of good quality and useful navigation profiles from these patterns. The discovery of aggregate usage profiles or patterns through clustering, as well as other Web mining techniques, have been explored by several research groups. However, in all of these cases, the frameworks proposed have not been extended to show how these profiles can be used as a part of a recommender system [9]. Normally, users cannot be identified from anonymous web server logs due to many reasons like using of a single computer by multiple users, and dynamic IPs. These anonymous users that every company wishes to attract and get involved with its website when they first visit the site [17] which means only the current active navigation is needed rather than a full user profile. As mentioned before, most of the previous researches are user-centered recommendation systems. We used the clusters produced by the clustering step (previous step) to build the usage or the navigation profile with one profile for each cluster. The navigation profile contains only those web pages that passed certain confidence support and weights values. The confidence support determines the frequency occurrence on those pages in the cluster. We stress again that these profiles ISBN:

4 don t consider specific users since we don t take the users history in account during obtaining the profiles. To summarize we construct a navigation profile as a set of pageview-weight pairs: profile = { p, weight(p) p Є P, weight(p) min_weight } where P ={p1, p2,..., pn}, a set of n pageviews appearing in the transaction file with each pageview uniquely represented by its associated URL and the weight(p) is the (mean) value of the attribute s weights in the cluster. Fig.3 shows a navigation profile Database snapshot of two profiles obtained for two clusters 1 and 2 where each profile contains related page views. For example, profile 1 represents the activity of a user interested in the courses and the programs offered while profile 2 represents the activity of a user interested in the pageviews related to the admission and advising. Fig. 3 Examples of Navigation Profiles Database 2) Online Component A. Experimental Setup Based on the proposed Architecture, a recommendation system is developed using Microsoft VC++ connected to Microsoft Access database through an Open Database Connection (ODBC). We used CTI dataset which contains sessions with 683 pageviews for the experiments with 70% for training and 30% for testing. It contains four files. Cti.cod, cti.tra, cti.std and cti.nav. as mentioned before, we used the cti.std file to generate the clusters and building the navigation profiles. The session forms the rows of this file and the pageviews form the columns. For the testing purpose, we used cti.tra which contains the filtered sessionzed data in transaction format. Each line in this file corresponds to the sequence of pages visited during one session. There is a direct correspondence between the rows in this file and the rows in the file "cti.std". While the order of occurrence of pageview in each session represents the order in which these pageviews were visited, the transactions do not contain repeated visited to the same pageview in the same session. Thus, only the first access to a pageview is recorded as part of the transaction. We used the precision, coverage and F1 In order to evaluate the recommendation effectiveness. Assume that we have active current session A taken from the evaluation set and we have R as a recommendation set using the prediction engine over the navigation profiles. W represents the items that already visited by the user in A. The precision is defined as: R ( A w) Pr ecision( = R (2) and the coverage is defined as: When the user navigates the internet, the web server will start to keep his logs on a file. This file can be accessed to extract the current active navigation web pages called the Active Session. Using this active session, the online component is responsible for assigning this user activities to the best navigation profile where by a recommendation list is to be created and attached to the user navigation list. We used a statistical analysis for the matching purpose. Since the active session and the choose profile can be represented as vectors; the cosine coefficient commonly used in information retrieval was used to do the matching purpose [9]. A recommendation score is computed for those items not already visited by the user in the active session in order to recommend them based on their scores. Re cscore = [ PageWeight* Pr ofilematch] Two factors are used in determining this recommendation score: the overall similarity of the active session to the profile as a whole, and the average weight of each item in the profile computed during obtaining the profiles (offline component). IV. RESULTS AND DISCUSSION (1) R ( A w) coverage( = (3) ( A w) Finally, F1 is defined as F1( 2 Pr ecision( cov erage( Pr ecision( + coverage( = (4) B. Evaluation of the recommendation Accuracy Table I shows the recommendation set contain pages recommended from the profile No.2 to the session with the following pages (/news/default.asp, /courses/, /courses/syllabisearch.asp, /courses/syllabilist.asp, /people/facultyinfo.asp?id=231) each with a recommendation score. That s mean the recommendation systems has chooses the profile No.3 (programs and advising) as the best source for predicting the next visited pages since this session contains pages pertaining courses. TABLE I EXAMPLE OF RECOMMENDATION SET FOR ANONYMOUS USER S SESSION Recommende d Profile 2 Recommendation Set /admissions/ /cti/advising/display.asp?page=intranetnew s Recommendatio n Score ISBN:

5 /cti/advising/login.asp /courses/syllabilist.asp /courses/ /people/ Table II shows the precision accuracy values obtained for different K-clusters. The best accuracy is gained when K is equal to 10. It shows that different K-clusters gives different precision. This is true because the clustering size does matter for calculating the confidence support and the mean weights. TABLE II PRECISION S SAMPLE FOR DIFFERENT K-CLUSTERS k Recommendation Precision Since the current user is anonymous to the recommender system with no previous navigation history, hence a sliding window technique over the current user session was used to represent the user history. To do so, the user current session is broken into two parts; the first part with size n pages is used as the surrogate user history which is matched against the web navigation profiles then produces a recommendation list from the selected profile. The remaining pages form the second part which is used for the comparison purpose to evaluate the recommendation accuracy. Since the average size of the dataset used for the experiments is pages per session, we used a window size equal to 2 to represent the surrogate user history and the rest pages are uused for the evaluation purpose. Fig.4 relates the recommendation effectiveness for our system (Anonymous user Recommender System ARS) with window size =2 pages compared to the findings of a previous methods PACT [9]. With a recommendation score s threshold varies from 0.1 to 1.0, the F1 measurement as a performance evaluation shows that ARS performs better and achieves higher prediction accuracy. This improvement is due to the mining processes applied to the extracted navigation profiles by the offline component followed by a better classification of the current user to the best web navigation profiles. Fig. 4 F1 recommendation measurement with window size=2 V. CONCLUSION AND FUTURE WORKS Web is one of the main sources of the information. The ability of predicting the next visited pages and recommending it to the user is highly recommended specially in e-commerce applications. We proposed ARS recommender system to predict the anonymous user next navigation based on preformed navigation profiles for similar interested users. The results showed higher prediction accuracy. The results also proved that web usage mining has the ability to build recommendation system for the anonymous user based on their early navigation which is applicable for e-commerce companies which like to turn users into customers. Rebuilding the profiles is a time consuming process. Adaptive profiles are one of the future interests. REFERENCES 1. Markov, Z. and D. Larose, Data Mining the Web: Uncovering Patterns in Web Content, Structure, and Usage. 2007: Wiley- Interscience. 2. Liu, B., ed. Web Data Mining: Exploring Hyperlinks, Contents, and Usage Data Data-Centric Systems and Applications. 2007, Springer. 3. Yang, Q. and X. Wu, 10 Challenging Problems In Data Mining Research. International Journal of Information Technology & Decision Making, (4): p Srivastava, J., et al., Web usage mining: discovery and applications of usage patterns from Web data. SIGKDD Explorations, (2): p Jalali, M., et al., WebPUM: A Web-based recommendation system to predict user future movements. Expert Systems with Applications, (9): p Yan, T.W., et al., From user access patterns to dynamic hypertext linking. Computer Networks and ISDN Systems, (7-11): p Mobasher, B., R. Cooley, and J. Srivastava, Automatic personalization based on web usage mining. Communications of the ACM, (8): p Nakagawa, M. and B. Mobasher, A hybrid web personalization model based on site connectivity, in Proceedings of the 5th International WebKDD. 2003, Citeseer. p Mobasher, B., et al., Discovery and Evaluation of Aggregate Usage Profiles for Web Personalization. Data Mining and Knowledge Discovery, (1): p Zhou, B., S.C. Hui, and K. Chang. An intelligent recommender system using sequential web access patterns. in 2004 IEEE Conference on Cybernetics and Intelligent Systems Liu, H. and V. Kešelj, Combined mining of Web server logs and web contents for classifying user navigation patterns and predicting users' future requests. Data & Knowledge Engineering, (2): p Baraglia, R. and F. Silvestri, An online recommender system for large Web sites, in Proceedings of the 2004 IEEE/WIC/ACM International Conference on Web Intelligence 2004, IEEE Computer Society: China. p Baraglia, R. and F. Silvestri, Dynamic personalization of web sites without user intervention. Communications of the ACM, (2): p Cornelis, C., et al., One-and-only item recommendation with fuzzy logic techniques. Information Sciences, (22): p Castellano, G., A.M. Fanelli, and M.A. Torsello, NEWER: A system for NEuro-fuzzy WEb Recommendation. Applied Soft Computing, (1): p ISBN:

6 16. Fayyad, U.M., G. Piatetsky-Shapiro, and P. Smyth, From Data Mining to Knowledge Discovery: An Overview, in Advances in knowledge discovery and data mining. 1996, American Association for Artificial Intelligence AAAI. p Suryavanshi, B.S., N. Shiri, and S.P. Mudur, Adaptive Web Usage Profiling, in Advances in Web Mining and Web Usage Analysis. 2006, Springer Berlin Heidelberg. p ISBN: