A Social Recommendation Model based on dual attention mechanism

Abstract

Social recommendation can effectively alleviate the information overload problem faced by users in social platforms. However, in the existing recommendation methods, problems such as sparse data and low recommendation accuracy do exist. Previous studies have paid more attention to explicit feedback data such as review ratings, and often ignored the influence of implicit feedback data such as social trust relationship on user interest preferences. In response to these problems, this paper first proposes Social Relationship Recommendation Model with Multi-level attention (MA_SRec), to solve the problem that deep and shallow social trust relationships have variant degree of impacts on users' interest preferences. The model uses LSTM Long Short-Term Memory network to extract user dynamic trust network, learns the influence weight of recent social trust relationship through attention mechanism. Assigning respective influence to features can accurately describe the user's interest preferences. This paper draws on the real information and social relationship data set Epinions, and compares MA_SRec with the social algorithm SErec, SBPR, implicit feedback recommendation algorithm WRMF and other classic algorithms. The four indicators of MA_SRec have improved significantly. The experimental results show that the method can deeply analyze social information and effectively improve the problem of data sparsity. This model supports application in information recommendation and various social recommendation scenarios.