ICS Ph.D. Defense (Pei-Chia Chang)

September 3, 10:30am - 12:00am
Mānoa Campus, POST 302

Title: A Personalized Recommender Agent For The World Wide Web – A Semantic Perspective

Abstract:
Web personalization aims to provide useful Internet information and avoid information overload. Most web personalization studies focus on a single website and recommending pages across websites is challenging due to a variety of concerns, such as dynamic web and diverse interests, conflict of interest, shared communication protocols required, and model reusability.

This work explores the potential of augmenting Wikipedia’s categories with page keywords for semantic user modeling to recommend pages across websites. Our recommender agent focuses on modeling individual web users’ topical interests, using the content-based usage analysis at the client-side. Our system also promotes serendipity as a major factor in our recommendations by considering the coverage of a user’s interests via the Diversity Index using the categorical topology.

We evaluated the system’s performance on recommendations regarding topicality and serendipity with 25 participants in the computer science domain. Results indicate that our system’s performance is significantly better than the pure content-based vector space model, particularly regarding serendipity. This is possibly due to the augmentation of Wikipedia’s categories with keywords as well as the utilization of categories’ topology.

This work is significant for four reasons. First, we emphasize the convergence between content modeling and user modeling by means of augmenting Wikipedia’s content and usage mining. Second, using the semantics (vocabulary, categorical association, etc.) of Wikipedia for user modeling is a worthwhile attempt. Our model is deliberately constructed as a research platform based on heuristic information extraction on keywords and allows for more heuristics. Third, we model a user’s topical interests using Wikipedia’s categories, which yield a simple model that can be interoperated among different websites. The model is at the client side, allowing more user control and reducing privacy concerns. Fourth, a methodology is supplied to researchers for further development of similar recommender agents.


Event Sponsor
Information and Computer Sciences, Mānoa Campus

More Information
Henri Casanova, 956-2649

Share by email