Context-aware graph embedding with gate and attention for session-based recommendation

Abstract

Prior solutions on session-based recommendation (SBR) are mainly limited by two major issues: (1) the sequence and transition relationships of items need further integration; (2) the context clues from the neighboring sessions remain largely under-explored. To overcome these issues, this paper proposes a novel Context-aware Graph Embedding Network (CGENet) with gate and attention mechanisms, that not only can effectively exploit the collaborative relationship between the sequence and transition patterns in each session, but also benefit greatly from topological context patterns among different sessions. Specifically, the proposed CGENet consists of three different parts, i.e., Transition Pattern Learning (TPL) module, Sequential Pattern Learning (SPL) module, and Context Pattern Learning (CPL) module. The TPL module is built on Graph Isomorphic Network (GIN) with multiple information highways to capture the transition relationships between items. To maximize the value of sequence-position information, a Gated Multilayer Perceptron (gMLP) is introduced into the SPL module to model the long-term dependencies between sequence tokens. Under the standardized guidance of the graph attention layer, the CPL module can further explore the topological contexts from neighboring sessions, thereby enhancing its ability to predict user preferences more effectively. Extensive experiments on three benchmark datasets demonstrate the superiority of the proposed CGENet compared to the state-of-the-art baselines.