基于多关系感知和自监督学习的超图增强会话推荐模型

A Hypergraph-Enhanced Session Recommendation Model Based on Multi-Relationship Awareness and Self-Supervised Learning

摘要: 会话推荐旨在基于匿名用户的历史行为序列预测下一个最有可能交互的项目。现有的大多数基于超图解决会话推荐模型没有考虑超边同构，忽略了会话内可能存在的多种依赖关系，从而无法捕获会话的潜在信息。同时，会话匿名性和短序列往往会导致数据稀疏性，影响了推荐模型的性能。因此本文提出一种基于多关系感知和自监督学习的超图增强会话推荐模型(MS-HGNN)，该模型能够有效捕获会话间用户的复杂高阶关系和会话内顺序及共现依赖关系。具体来说，MS-HGNN模型将会话序列分别建模为超图、共现图和顺序图，在捕获超图的同时加入位置信息以减轻超边同构，利用门控机制融合会话内顺序依赖关系和共现依赖关系。此外，使用自监督学习最大化两个会话表示之间的互信息以缓解数据稀疏性，有效地融合用户的复杂高阶关系和会话内顺序及共现依赖关系。广泛的实验结果表明，所提出的MS-HGNN模型在三个真实数据集Diginetica、Tmall和Last.fm上的性能优于目前最先进的模型。

Abstract: Session-based recommendation aims to predict the next most likely item for interaction based on the historical behavior sequence of anonymous users. Most existing hypergraph-based solutions for session recommendation overlook hyperedge isomorphism and ignore the various dependencies that may exist within a session, thus failing to capture the latent information of the session. Meanwhile, the anonymity of sessions and the short sequences often lead to data sparsity, affecting the performance of recommendation models. Therefore, this paper proposes a Multi-Rela- tionship Aware and Self-Supervised Learning-based Hypergraph-Enhanced Session-based Recommendation Model (MS-HGNN), which effectively captures the complex high-order relationships among users across sessions and the sequential and co-occurrence dependencies within sessions. Specifically, the MS-HGNN model models session sequences as hypergraphs, co-occurrence graphs, and sequential graphs, respectively. It incorporates positional information to mitigate hyperedge isomorphism while capturing hypergraphs, and utilizes a gating mechanism to fuse the sequential and co-occurrence dependencies within sessions. Additionally, self-supervised learning is employed to maximize the mutual information between the representations of two sessions, alleviating data sparsity and effectively integrating the complex high-order relationships among users and the sequential and co-occurrence dependencies within sessions. Extensive experimental results demonstrate that the proposed MS-HGNN model outperforms the current state-of-the-art models on three real-world datasets: Diginetica, Tmall, and Last.fm.