Behavioral similarities for collaborative recommendations

Recommender systems contribute to the personalization of resources on web sites and information retrieval systems. These systems use different mining techniques to generate predictions corresponding to the needs of users. In this paper, we present a hybrid recommender system using a userbased approach, that combines predictions based on web usage patterns and rating data. We suggest a new technique that takes into account common usage patterns in order to compute correlations between users and select neighborhoods, without using any rating data. Unlike classical predictive systems that use directly mining techniques in order to suggest recommendations (like the Longest Common Subsequences technique or the frequent sequential patterns mining), the originality of our usage based technique is not to harness the discovered patterns to recommend resources, but to assess similarities of navigational users profiles. Recommendations are performed in a following step. The performance of our system is tested without and by combining predictions of both navigational based technique and classical collaborative filtering, in terms of accuracy and robustness. The experimentation put forward the impact of the navigational based technique on the performance of the hybrid recommender system in terms of precision and robustness. The tests show that the more the navigational based technique is involved in the recommendation process, the more the best predictions are accurate.