A Hybrid Approach for Intelligent Recommender Systems

Abstract

The anticipation of the user's next move is one the main techniques needed for web personalization. Next page prediction aims at discovering the next page the user will visit for offering recommendations as well as pre-fetching to reduce network latency. In this work we proposed a next page prediction system that is based on a hybrid framework combining memory-based and model based recommender systems. We offered three different representations of user preferences and tested their results on different datasets. The first representation reflected usage data by building a user-page matrix. The second approach incorporated semantic information to build a user-concept matrix. Finally the third approach offered two methods to combine usage and semantic data. The approaches yielded a 12.8% and 33% improvement in prediction accuracy for the first two approaches respectively, and a 47.3% and 54.3% for the combined approach. The system also used clustering to group users and frequent patterns which caused the prediction time to be reduced by an average of 69.2%.