Incremental Relational Fuzzy Subtractive Clustering for Dynamic Web Usage Profiling

A primary application of web usage profiling is in model-based collaborative filtering (CF) for building recommender systems used for web personalization. CF techniques used for recommendation require accumulation of vast amount of historical user-preference information, which is queried to provide a personalized experience. Model-based CF techniques are preferred over the somewhat more accurate memory-based CF techniques primarily due to their higher efficiency and scalability. In the situation where user interests change dynamically with time, memory-based CF allows easy addition of new usage data, but model-based CF often requires complete remodeling operation. Being computationally intensive, remodeling is done occasionally and the models used normally lag behind the current usage patterns leading to irrelevant and mistargeted recommendations. Even though development of maintenance schemes which adapt these models to non-stationary environments is very important, it has received less attention so far. Our first contribution in this paper is a web usage profile maintenance scheme using a new algorithm called incremental Relational Fuzzy Subtractive Clustering (RFSC). Incremental RFSC can efficiently add new usage data to an existing model overcoming the expense associated with frequent remodeling. We validate the results by showing close similarity between complete reclustering and the clustering models obtained after applying our incremental RFSC technique. Any maintenance scheme based on incremental update of the profile requires a measure to indicate, to the web analyst, when accumulated usage data has to be reclustered; otherwise continued maintenance leads to irrelevant, obsolete model. The second contribution of this paper is thus introduction of a quantitative measure, called impact factor. When the impact factor exceeds a predefined threshold, a remodeling is recommended. We conducted extensive experiments which compare the effectiveness of recommendations using our incremental RFSC technique, complete reclustering, and memory-based CF techniques. The results obtained indicate that our maintenance technique is almost as good as complete remodeling.