Multi-behavior Recommendation Based on Simplified Graph Convolutional Networks

There are a large amount of user behavior data in the recommendation system. The types of user behaviors are various, such as click, favorite, like, purchase and so on. However, most of the existing recommendation methods only consider single user behavior, such as purchase, ignoring the influence of auxiliary behavior data on target behavior. In order to make full use of multi-behavior data and further improve the recommendation performance, multi-behavior recommendation based on simplified graph convolutional networks(SGCNMB) is proposed in this paper. The SGCNMB model firstly initializes the embedding vectors for users and items in the shared embedding layer; then, in the embedding propagation layer, the simplified graph convolutional network is exploited to learn the user representations and item representations on a single behavior, the users' preferences on different single behavior are calculated respectively; finally, through the fusion prediction layer, the users' preferences on various behaviors are fused to get the top-n recommendation results of users on the target behavior. Extensive experiments were conducted on two real-world multi-behavior datasets. The experimental results demonstrate that SGCNMB can effectively take advantage of multi-behavior interaction data to further improve the recommendation performance.