面向会话表示的多阶段自监督集合学习方法

Multi-Stage Self-Supervised Set Learning Approach for Session Representation

摘要: 会话推荐旨在根据用户在会话单元中的交互组对用户下一项感兴趣的交互进行推荐。然而，现有会话推荐方法通常将会话建模为序列形式，使推荐结果过于依赖交互顺序。为了应对这一问题，本文提出了一种面向会话表示的多阶段自监督集合学习方法SR-MSSL。该方法将会话建模为集合形式，通过自注意力机制对集合中的交互项进行信息聚合，并通过多阶段自监督学习挖掘元素间的隐含关联，从而在不依赖交互顺序的情况下获取具有表达力的会话表示。实验在Yoochoose和Diginetica数据集上进行，结果显示，本文方法在多个指标上优于传统推荐方法与多种深度学习模型，验证了集合建模在会话表示学习中的有效性。此外，通过位置编码消融实验进一步证明，去除交互的顺序信息能够提高会话模型的推荐性能。

Abstract: Session-based recommendation seeks to predict the next interaction of interest for a user by leveraging the set of interactions within a session. However, current session-based recommendation approaches usually treat sessions as sequences, making the recommendations excessively dependent on the order of interactions. To mitigate this limitation, we propose SR-MSSL, a multi-stage self-supervised set learning framework for session representation. The approach models sessions as sets, employing a self-attention mechanism to aggregate information across interactions and leveraging multi-stage self-supervised learning to capture latent relationships among elements, thus deriving expressive session representations independent of interaction order. Experiments on the Yoochoose and Diginetica datasets demonstrate that our method surpasses traditional recommendation techniques and several deep learning models across multiple evaluation metrics, confirming the effectiveness of set-based modeling for session representation learning. Additionally, ablation studies on positional encoding reveal that discarding the sequential information of interactions can further improve the recommendation performance of session models.