Combining trust in collaborative filtering to mitigate data sparsity and cold-start problems

Abstract

Collaborative filtering (CF) is the most popular approach to build recommender systems and has been successfully employed in many applications. However, it suffers from several inherent deficiencies such as data sparsity and cold start. To better show user preferences for the cold users additional information (e.g., trust) is often applied. We describe the stages based on which the ratings of an active user's trusted neighbors are incorporated to complement and represent the preferences of the active user. First, by discriminating between different users, we calculate the significance of each user to make recommendations. Then the trusted neighbors of the active user are identified and aggregated. Hence, a new rating profile can be formed to represent the preferences of the active user. In the next stage, similar users probed based on the new rating profile. Finally, recommendations are generated in the same way as the conventional CF with the difference that if a similar neighbor had not rated the target item, we will predict the value of the target item for this similar neighbor by using the ratings of her directly trusted neighbors and applying MoleTrust algorithm, so as to incorporate more similar users to generate prediction for this target item. Experimental results demonstrate that our method outperforms other counterparts both in terms of accuracy and coverage.