A lightweight graph convolutional network-based session recommendation method based on multihop attention mechanism

Abstract

Conversational recommendation aims to predict a user's next potentially interesting item based on a series of historical clicks generated in an anonymous session. In recent years, Graph Convolutional Networks (GCNs) have achieved great success in this field due to their excellent ability to capture complex item transformation relationships. However, existing GCN-based recommendation models suffer from three main limitations: traditional GCN layers contain dense feature transformation matrices and nonlinear activation functions, which not only incur high computational costs in highly sparse and low-feature-dimensional scenarios like conversation graphs but also easily lead to overfitting. As the number of network layers increases, node representations inevitably tend to homogenize (i.e., oversmoothing). Finally, most graph recommendation models are mainly limited to local aggregation of single-hop neighbors, making it difficult to effectively capture the non-local, multi-hop jump dependencies prevalent in conversation sequences. To address these challenges, this paper proposes a novel Multi-Hop Lightweight Graph Convolutional Network (MH-LGCN) conversational recommendation model based on a multi-hop attention mechanism. Specifically, MH-LGCN first abandons redundant operations in traditional graph convolution and designs a lightweight information propagation mechanism that retains only neighbor feature aggregation. This not only greatly improves the training efficiency of the model but also fundamentally alleviates the oversmoothing phenomenon caused by deep propagation. Simultaneously, this paper designs a hierarchical multi-hop attention mechanism that adaptively captures the user's long-term click preferences and skipping intentions by calculating the attention weights between the current interactive item and multi-level context items. Extensive experiments were conducted on three publicly available real-world datasets: Diginetica, Tmall, and Yoochoose 1/64, covering comparisons with 10 state-of-the-art baseline models. Experimental results show that MH-LGCN achieves state-of-the-art performance on both Precision@20 and MRR@20, outperforming the existing strong baseline model GCE-GNN by up to 4.1% and 5.8%, respectively. Furthermore, detailed ablation experiments, parameter sensitivity analysis, and model complexity comparisons further validate the effectiveness and efficiency of the proposed modules.