A collaborative filtering approach to predict web pages of interest from navigation patterns of past users within an academic website.

摘要

This dissertation is a simulation study of factors and techniques involved in designing hyperlink recommender systems that recommend to users, web pages that past users with similar navigation behaviors found interesting. The methodology involves identification of pertinent factors or techniques, and for each one, addresses the following questions: (a) room for improvement; (b) better approach, if any; and (c) performance characteristics of the technique in environments that hyperlink recommender systems operate in. The following four problems are addressed:; Web page classification. A new metric (PageRank x Inverse Links-to-Word count ratio) is proposed for classifying web pages as content or navigation, to help in the discovery of user navigation behaviors from web user access logs. Results of a small user study suggest that this metric leads to desirable results.; Data mining. A new apriori algorithm for mining association rules from large databases is proposed. The new algorithm addresses the problem of scaling of the classical apriori algorithm by eliminating an expensive join step, and applying the apriori property to every row of the database. In this study, association rules show the correlation relationships between user navigation behaviors and web pages they find interesting. The new algorithm has better space complexity than the classical one, and better time efficiency under some conditions and comparable time efficiency under other conditions.; Prediction models for user interests. We demonstrate that association rules that show the correlation relationships between user navigation patterns and web pages they find interesting can be transformed into collaborative filtering data. We investigate collaborative filtering prediction models based on two approaches for computating prediction scores: using simple averages and weighted averages. Our findings suggest that the weighted averages scheme more accurately computes predictions of user interests than the simple averages scheme does.; Clustering. Clustering techniques are frequently applied in the design of personalization systems. We studied the performance of the CLARANS clustering algorithm in high dimensional space in relation to the PAM and CLARA clustering algorithms. While CLARA had the best time performance, CLARANS resulted in clusters with the lowest intra-cluster dissimilarities, and so was most effective in this regard.