Hybrid Sampling Light Graph Collaborative Filtering for Social Recommendation

Abstract

• The use of graph neural networks has been widely adopted in recommender systems as a state-of-the-art collaborative filtering mechanism. In graph neural collaborative filtering, extracting negative signals from implicit feedback aris-ing from the interaction between users and items is a ma-jor challenge. The negative sampling aspect has not been fully explored in the use of graph neural collaborative filtering for the social recommendation. This study explores negative sampling by combining a graph neural network aggregation procedure with social recommendation graph structures. A system called Hybrid Sampling Light Graph Convolution Collaborative Filtering for Social Recommendations (HLCS) is proposed in this paper. Through the propagation and fusion of embedded representations of users and items in the item domain and social domain, hard negative samples are generated by the hybrid sampling technique to optimize the recommendation model’s performance. Using two real-world datasets, we conducted comprehensive experiments and showed that the HLCS approach was superior to the SOTA approach, particularly in cold-start situations. ;