Recommender systems (RS) aim to suggest relevant products such as items, films, and music based on user preferences, and have long been a focus in both industry and academia. While traditional RS approaches, including collaborative filtering (CF) and content-based methods, have achieved considerable success, they often struggle with challenges such as data sparsity and limited modeling of complex relations. Recently, graph neural networks (GNNs) have emerged as powerful tools for RS by capturing rich structural and semantic dependencies in user-item interactions. This paper presents a comprehensive survey of GNN-based RS with a particular focus on multi-type scenarios, where multiple behaviors, data modalities, or graph relations coexist. We first introduce a clear taxonomy that distinguishes multi-types in RS and classify existing GNN-based models according to graph construction strategies, embedding learning mechanisms, and application contexts. We then analyze the motivations for adopting GNNs in multi-type RS and review representative methods, highlighting how they address challenges in framework design, embedding layer construction, and computational efficiency. Finally, we discuss open issues such as scalability, interpretability, and dynamic graph modeling, and outline potential research directions to guide future advancements in this field.
Liao et al. (Fri,) studied this question.