区分邻居角色，让多模态图神经网络更精准地传播信息。
RoleMAG: Learning Neighbor Roles in Multimodal Graphs
- 为邻居分配共享、互补或异质角色，分通道传播信号。
- 在RedditS和Bili_Dance上超越现有方法，提升多模态节点分类性能。
- 适合需要精细建模多模态关系的图学习任务，如社交推荐、视频分析。
多模态属性图（MAGs）结合了多模态节点属性与结构关系。然而，现有方法通常在单一图上进行共享消息传递，隐含假设所有模态的邻居同样有用。实际上，对某一模态有益的邻居可能干扰其他模态，导致模态特异性信号被模糊。为此，我们提出RoleMAG，一种学习邻居角色的多模态图框架。具体而言，RoleMAG区分邻居应提供共享、互补或异质信号，并通过独立传播通道处理。这使得互补邻居能实现跨模态补全，同时将异质邻居排除在共享平滑之外。在三个以图为中心的MAG基准上，RoleMAG在RedditS和Bili_Dance上取得最佳表现，且在Toys上保持竞争力。消融实验、鲁棒性分析和效率评估进一步验证了所提角色感知传播设计的有效性。代码已开源。
原文摘要 · Abstract (English)
Multimodal attributed graphs (MAGs) combine multimodal node attributes with structured relations. However, existing methods usually perform shared message passing on a single graph and implicitly assume that the same neighbors are equally useful for all modalities. In practice, neighbors that benefit one modality may interfere with another, blurring modality-specific signals under shared propagation. To address this issue, we propose RoleMAG, a multimodal graph framework that learns how different neighbors should participate in propagation. Concretely, RoleMAG distinguishes whether a neighbor should provide shared, complementary, or heterophilous signals, and routes them through separate propagation channels. This enables cross-modal completion from complementary neighbors while keeping heterophilous ones out of shared smoothing. Extensive experiments on three graph-centric MAG benchmarks show that RoleMAG achieves the best results on RedditS and Bili\_Dance, while remaining competitive on Toys. Ablation, robustness, and efficiency analyses further support the effectiveness of the proposed role-aware propagation design. Our code is available at https://anonymous.4open.science/r/RoleMAG-7EE0/
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。