Social influence-weighted user preference for recommendation systems

Social recommendation significantly enhances the performance of recommendation systems by introducing social networks, which play a crucial role in current online social platforms. The foundation of social recommendation is the social influence hypothesis, which posits that social relationships can influence user preferences. This implies that social relationships and user preferences relationships are distinct yet related. However, most current studies oversimplify social relationships as ”biased” user preference relationships and subsequently focus on eliminating ”relationship bias”. This leads to information loss and significantly limits the expressive power of social networks. To address this challenge, we propose Social Influence-weighted User Preference (SIUP). Our approach introduces a novel perspective to reassess the effect of social networks on recommendation systems. In this perspective, social networks are leveraged to extract social influence, which is then utilized to quantify the weight of friends on user preferences. It fully exploits the information from social networks and genuinely adheres to the social influence hypothesis. Additionally, SIUP constructs fine-grained preference influence factors and employs attention-based neighborhood information aggregation to obtain latent preference distributions of users for items. We conducted experiments on three publicly available datasets (Ciao, Yelp, and Epinions), each consisting of tens of thousands of nodes. The experimental results show that SIUP obtains an average performance improvement of 3.24% in Recall and 8.18% in NDCG. The code is available at https://github.com/mgf9505/SIUP-main.