Assessing usage of negative similarity and distrust information in CF-based recommender system

Nowadays recommender systems are developed to provide information for users to choose the best things that they want. Recently different collaborative filtering techniques have been successfully employed to provide precise recommendations. However, sparsity problem is still considered as an important remained challenges. Existing CF based recommendation methods generally focus on positive similarities and trust. In this approach the majority of data containing negative, zero similarity and distrust, are omitted during prediction; which aggravates sparsity problem further. For each target user, we propose to partition the set of the other users into distinct groups according to the state of similarity and trust information. We consider positive zero and negative similarities as well as trust, distrust and zero trust information. In this paper we present a new concept to show that all data in recommender systems are important and can help us to predict new items better. We divide dataset information to nine partition and calculate each partition effectiveness to predict new item separately.