Analyzing Recommender System's Performance Fluctuations across Users

Recommender systems (RS) are designed to assist users by recommending them items they should appreciate. User based RS exploit users behavior to generate recommendations. As a matter of fact, RS performance fluctuates across users. We are interested in analyzing the characteristics and behavior that make a user receives more accurate/inaccurate recommendations than another. We use a hybrid model of collaborative filtering and trust-aware recommenders. This model exploits user’s preferences (represented by both item ratings and trusting other users) to generate its recommendations. Intuitively, the performance of this model is influenced by the number of preferences the user expresses. In this work we focus on other characteristics of user’s preferences than the number. Concerning item ratings, we touch on the rated items popularity, and the difference between the attributed rate and the item’s average rate. Concerning trust relationships, we touch on the reputation of the trusted users.