EFBH: Collaborative Filtering Model Based on Multi-Hypergraph Encoder
Zhe Yang, Liangkui Xu, Lei Zhao
Abstract
Recommendation systems are necessary to enhance the consumer shopping experience. As one of the classical algorithms, collaborative filtering (CF) is the main recommendation model. Currently, the graph neural network realize collaborative filtering through graph convolution and model the interaction between users and items. However, the ability of ordinary graph to express the higher-order relationship between vertices and to process sparse data is limited. It cannot fully capture the potential interaction between users and items. So the hypergraph is used to model the interaction, and the recommendation question is transformed into the link prediction problem on the hypergraph. An Encoder Framework based on Bipartite Hypergraph (EFBH) is proposed. First, the hypergraph convolution module is constructed, and the user-user hypergraph and the item-item hypergraph are constructed to solve the data sparsity problem; then the multi-hypergraph encoder module is constructed to encode the cooperation signals hidden in the user-item interaction to obtain high-quality vertex embedding vectors and better recommendation results. Experiments on three common datasets show that our model improves the AUC on average by 5.35%, 2.91%, and 3.26% respectively. The RMSE also decreased by an average of 1.43%, 15.33%, and 16.06% respectively, which is the best one compared to the existing models.