IASR: An Item-Level Attentive Social Recommendation Model for Personalized Ranking.

Most recommender systems provide recommendations by listing the most relevant items to a user. Such recommendation task can be viewed as a personalized ranking problem. Previous works have found it advantageous to improve recommendation performance by incorporating social information. However, most of them have two primary defects. First, in order to model interaction between users and items, existing works still resort to biased inner production, which has proved less expressive than neural architectures. Second, they do not delicately allocate weights of social neighbor influence based on the user feature or the item feature in a recommendation task. To address the issues, we propose an Item-level Attentive Social Recommendation model, IASR for short, in this paper. It employs an item-level attention mechanism to adaptively allocate social influences among trustees in the social network and gives more accurate predictions with a neural collaborative filtering framework. Extensive experiments on three real-world datasets are conducted to show our proposed IASR method out-performs the state-of-the-art baselines. Additionally, our method shows effectiveness in the cold-start scenario.