DH-HGCN: Dual Homogeneity Hypergraph Convolutional Network for Multiple Social Recommendations

Abstract

Social relations are often used as auxiliary information to improve recommendations. In the real-world, social relations among users are complex and diverse. However, most existing recommendation methods assume only single social relation (i.e., exploit pairwise relations to mine user preferences), ignoring the impact of multifaceted social relations on user preferences (i.e., high order complexity of user relations). Moreover, an observing fact is that similar items always have similar attractiveness when exposed to users, indicating a potential connection among the static attributes of items. Here, we advocate modeling the dual homogeneity from social relations and item connections by hypergraph convolution networks, named DH-HGCN, to obtain high-order correlations among users and items. Specifically, we use sentiment analysis to extract comment relation and use the k-means clustering to construct item-item correlations, and we then optimize those heterogeneous graphs in a unified framework. Extensive experiments on two real-world datasets demonstrate the effectiveness of our model.