Clustering Users to Explain Recommender Systems' Performance Fluctuation

Shareef Haydar (1) , Azim Roussanaly (1) , Anne Boyer (1)

1 KIWI - Knowledge Information and Web Intelligence

Recommender systems (RS) are designed to assist users by recommending them items they should appreciate. User based RS ex- ploits users behavior to generate recommendations. Users act in accor- dance with different modes when using RS, so RS's performance fluctu- ates across users, depending on their act mode. Act here includes quan- titative and qualitative features of user behavior. When RS is applied in an e-commerce dedicated social network, these features include but are not limited to: user's number of ratings, user's number of friends, the items he chooses to rate, the value of his ratings, and the reputation of his friends. This set of features can be considered as the user's profile. In this work, we cluster users according to their acting profiles, then we compare the performance of three different recommenders on each cluster, to explain RS's performance fluctuation across different users' acting modes.

Mots clés: Recommender system; collaborative filtering; trust-aware; trust; reputation; user profile; clustering; item popularity; abnormality

Domaines: Intelligence artificielle [cs.AI]; Recherche d'information [cs.IR]; Web

Shareef Haydar, Azim Roussanaly, Anne Boyer. Clustering Users to Explain Recommender Systems' Performance Fluctuation. ISMIS12 - The 20th International Symposium on Methodologies for Intelligent Systems - 2012, Dec 2012, Macau, China. pp.357-366, ⟨10.1007/978-3-642-34624-8_41⟩. ⟨hal-00777218⟩