基于集合表示学习的会话推荐

Session-Based Recommendation with Set Representation Learning

摘要: 会话推荐旨在预测用户在会话中的下一次交互。现有方法多依赖序列建模，假设交互存在严格的时间顺序，但实际场景中用户行为常呈现弱顺序或无序特征，易导致推荐不稳定。本文提出一种基于集合表示学习的会话推荐模型SetRec，将会话建模为无序集合，直接捕捉物品间的全局共现关系，降低对顺序噪声的依赖。为提升表征稳健性，模型引入自监督重构任务，通过随机遮蔽与恢复交互项，迫使编码器学习潜在依赖关系，并与推荐预测任务联合优化，实现“重构 + 预测”的协同增强。实验在Yoochoose与Diginetica两个公开数据集上进行，结果表明SetRec在准确率和排序质量上均优于主流方法，有效缓解了数据稀疏与顺序噪声带来的问题，验证了集合化建模与自监督机制在会话推荐中的优势。

Abstract: Session-based recommendation aims to predict the next interaction within a user session. Existing methods often rely on sequential modeling, assuming strict temporal dependencies among interactions. However, in real-world scenarios, user behaviors are usually weakly ordered or unordered, leading to unstable recommendations. This paper proposes SetRec, a session-based recommendation model based on set representation learning. By modeling sessions as unordered sets, SetRec directly captures global co-occurrence relationships among items and reduces sensitivity to sequential noise. To enhance representation robustness, the model introduces a self-supervised reconstruction task that randomly masks and recovers session items, forcing the encoder to learn latent dependencies. The reconstruction is jointly optimized with the recommendation prediction task, forming a “reconstruction + prediction” framework. Experiments on two public datasets, Yoochoose and Diginetica, demonstrate that SetRec outperforms mainstream methods in both accuracy and ranking quality, effectively mitigating the challenges of data sparsity and noisy order. These results verify the advantages of set-based modeling and self-supervised mechanisms for session-based recommendation.