A Graph Convolutional Network for Session Recommendation Model Based on Improved Transformer

Abstract

Graph convolutional networks are widely used for session-based recommendation (SBR) of products, aimed at solving anonymous sequence recommendation problems. However, currently almost all SBR models only focus on the current session, ignoring item transitions in other sessions. The paper introduces DA-GCN, a session-based recommendation model that utilizes graph convolutional networks. DA-GCN learns item embeddings from two perspectives, the global graph and the session graph: (1) The global graph updates the adjacency matrix through the shortest path algorithm, transforming the adjacency matrix from a single 0/1 information element to a complex dynamic graph with weights, and the global item embeddings are learned recursively through a session-aware attention mechanism; (2) The session graph learns session-level item embeddings by considering the item transitions in the current session graph and introduces an improved Transformer network when aggregating node information in the graph. The improved Transformer uses reverse position encoding to simulate the historical interests of the current session, while considering the correlation with global item embeddings. The DA-GCN model adopts an auxiliary loss function to supervise the historical interest extraction process, and then further models the correlation between the historical interests of the current session and the global item embeddings using attention mechanisms. The research uses three real-world datasets to demonstrate the effectiveness of the proposed method, and the results show an average improvement of 4.06% on the core metric P@20.