Abstract

Web personalization aims to provide content and services tailor-made to the needs of individual users usually from the knowledge gained through their (previous) interactions with the site. Typically, an access behavior model of users is learnt from the usage of the website which is then used to provide personalized recommendations to the current user(s). Clearly the performance of a recommender system will depend on the underlying model. Two fundamental challenges in personalization are effectiveness (accuracy) and efficiency (scalability) of recommender algorithms. In this thesis, we present a new class of techniques for efficient and effective web personalization. All the techniques are based on a new algorithm for fast mining of web usage data, called Relational Fuzzy Subtractive Clustering (RFSC). RFSC is scalable to large datasets, does not require user biased control parameters, is relatively more immune to noise which is inherent in usage data, and can capture overlapping user interest areas. Making innovative use of fuzzy grade of memberships and fuzzy prototypes of the access behavior model learnt through RFSC, we also propose improvements over cluster based and association rule based recommender models. Since browsing behavior of users on the web is not static but changes dynamically over time, we further propose a new maintenance scheme, which extends RFSC, to efficiently add new web usage data to an existing model in order to achieve adaptability over non-stationary volatile web environment. We validate our claims of effectiveness and efficiency through extensive experimentation on synthetic data as well as large datasets of web logs