A Hypergraph Augmented and Information Supplementary Network for Session-Based Recommendation

Jiahuan Chen, Caihong Mu, Mohammed Alloaa, Yi Liu

Abstract: Predicting the next interaction from a session without long-term historical data is challenging in session-based recommendation. Traditional approaches based on hypergraph modeling treat all items in sessions as interactions at the same time, producing hypergraphs that lose sequential information, which are susceptible to the interference of noise in long sessions. Besides, the way in which the information interacts through hyperedges in hypergraph convolution will lead to item embeddings lacking of global information. To address these issues, we propose a Hypergraph Augmented and Information Supplementary Network (HAISN), where the global graph self-supervised learning (GGSL) channel is designed to provide global and sequential information to the hypergraph. The Hypergraph augmented learning (HAL) channel is devised to supplement the hypergraph with filtered session information. Self-supervised learning is used to provide the information of the GGSL and HAL channels for the hypergraph convolution channel, improving the model effectively. Extensive experiments validate the effectiveness of HAISN.