An Anchor-Semantics-Aware Neural Preference Propagation Model for Session-based Recommendation.

Session-based recommendation aims to predict user's next action based on the anonymous sessions. Recent studies mainly apply Graph Neural Networks (GNN) to model the complex item-transitions and transfer the collaborative signals among sessions. However, most existing methods ignore the semantic inconsistency problem cause by item functional diversity and popular item during the process of cross-session collaborative signal capturing, which may lead to a preference negative transfer between irrelevant sessions. In addition, current works neglect to retain the semantic discriminability when learning the representations of sessions and items, and consequently degenerate the personalized recommendation. In this paper, we propose a novel model, called Anchor-Semantics-Aware Neural Preference Propagation (ASA-NPP), to simulate the transfer of collaborative signals with an anchor-semantics-aware recursive neural propagation over a designed graph. Specifically, we devise a Session Graph with Anchor Links (SGAL) based on all sessions, and then present an Anchor-Semantics Attention Network (ASAN) to transfer the semantically consistent cross-session collaborative signals and learn session-specific embeddings encoding different session semantics. Furthermore, we propose a Session Semantics Enhancement (SSE) module to improve the semantic discriminability of the learned representation via an elaborate self-supervised learning task. Extensive experiments on real-world datasets demonstrate the effectiveness and superiority of ASA-NPP over the state-of-the-art methods.