基于时序信息增强的图注意力网络会话推荐

Session-Based Recommendation Using Graph Attention Networks Enhanced with Temporal Information

摘要: 针对现有基于图神经网络的会话推荐模型过度侧重图结构空间关联、忽视会话时序特性，且会话编码器依赖单一项目嵌入、难以捕捉用户长短期真实兴趣及多阶段兴趣转移的问题，本文提出基于时序信息增强的图注意力网络会话推荐模型，首先通过时序信息增强模块，融合初始项目顺序嵌入与位置嵌入，借助多头自注意力强化时序全局依赖，并引入Dropout正则化与残差连接优化特征聚合；其次构建含自连接的有向会话图，利用多层图注意力网络自适应学习邻居节点权重，实现精准多跳信息传播与噪声抑制；最后设计多级意图会话编码器，按时间顺序从近到远划分多阶段兴趣，通过多头注意力学习各阶段贡献权重并聚合，避免单一编码的片面性。实验结果显示，该模型在Tmall与Diginetica两大电商数据集的P@10、P@20、MRR@10、MRR@20四项核心指标上，均显著优于包括Transformer基方法、最新GNN变体在内的主流基线模型。消融实验进一步验证时序信息增强模块、多层图注意力机制及多级意图编码器的关键有效性，并直接证明了多级意图编码器相较于标准的“序列最后隐藏层 + 全局注意力”机制的显著优势，证明模型能深度融合时序动态与图结构关联信息，显著提升会话推荐的精准度与合理性。

Abstract: Existing graph neural network-based session-based recommendation models often overemphasize graph structural spatial relationships while neglecting sequential characteristics of sessions. Furthermore, their session encoders rely on single item embeddings, making it difficult to capture users’ long-term and short-term interests and multi-stage interest shifts. To address these issues, this paper proposes a graph attention network-based session recommendation model enhanced with temporal information. First, a temporal information enhancement module fuses initial item sequential embeddings and positional embeddings, using multi-head self-attention to strengthen global temporal dependencies, and incorporating Dropout regularization and residual connections to optimize feature aggregation. Second, a directed session graph with self-connections is constructed, and a multi-layer graph attention network is used to adaptively learn neighbor node weights, enabling accurate multi-hop information propagation and noise suppression. Finally, a multi-level intent session encoder is designed to divide interests into multiple stages chronologically from near to far, learning the contribution weights of each stage through multi-head attention and aggregating them, avoiding the limitations of single encoding. Experimental results show that the proposed model significantly outperforms mainstream baseline models including Transformer-based methods and the latest GNN variants on four key metrics (P@10, P@20, MRR@10, and MRR@20) on two major e-commerce datasets, Tmall and Diginetica. Ablation experiments further validate the critical effectiveness of the temporal information enhancement module, multi-layer graph attention mechanism, and multi-level intent encoder, and directly prove the significant advantage of the multi-level intent encoder over the standard “last hidden layer of sequence + global attention” mechanism, demonstrating that the model can deeply integrate temporal dynamics and graph structural relational information, significantly improving the accuracy and rationality of session-based recommendations.