Hypergraph denoising neural network for session-based recommendation

Abstract

Session-based recommendation (SBR) predicts the next interaction of users based on their clicked items in a session. Previous studies have shown that hypergraphs are superior in capturing complex item transitions which contribute to SBR performance. However, existing hypergraph-based methods fail to model item co-occurrence and sequential patterns simultaneously, limiting the improvement of recommendation performance. Moreover, they are more sensitive to noisy items than conventional graph models due to the item association mechanism. In this paper, we propose a novel hypergraph-based method named Hypergraph Denoising Neural Network (HDNN) for SBR to tackle the abovementioned problems. The proposed method involves two newly-designed modules: a sequential pattern learning module (SPLM) and an adaptive attention selection module (AASM). In particular, SPLM models item sequential patterns to complement the hypergraph-based models which only focus on co-occurrence patterns. Meanwhile, AASM employs learnable attention score thresholds to exclude items with low attention scores, mitigating the impact of noisy items in hypergraphs. Furthermore, the sequential denoising unit (SDU) designed in SPLM is employed to eliminate noise in item sequential patterns, thus realizing the dual denoising purpose. Extensive experiments are conducted on three real-world datasets. The results of the experiments show that our HDNN framework shows better performance than the state-of-the-art models. In particular, all evaluation metrics in Tmall and RetailRocket showed improvements of over 15% and 5%, respectively.