.png){kind=icon}
.png){kind=icon}
Return
Graph Neural Networks With Interaction-Aware View Fusion for Graph Classification
.png){kind=icon}
PRE
AI
DOI：10.1109/TCSS.2025.3632799.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Graph neural networks (GNNs) have demonstrated outstanding performance in graph classification tasks. Most existing GNNs designed for graph classification adopt a structure that combines graph convolutional operators with graph pooling operators. However, these methods are often prone to being misled by spurious connections between nodes, which can arise due to noise or inherent data bias. Furthermore, these methods struggle to simultaneously capture the relationships between global and local information, leading to either overly smooth or overly sensitive classifications. To solve the above issues, we introduce a GNN with interaction-aware view fusion (IAVF) for graph classification, which effectively accounts for interactions between nodes to better handle both global and local information while mitigating the impact of misleading connections. Specifically, IAVF first generates multiple views using distinct strategies, ensuring that they can mutually correct each other during the subsequent fusion process. Then, we propose a global-local view interaction module, which captures the relational information between views at both global and local scales to achieve global-local interaction-aware ability. To ensure effective view fusion after interaction, we introduce a dual-supervision mechanism as a constraint to align features across views while preventing excessive alignment that may lead to an overly smooth model. On established and competitive benchmarks, even with only a single type of graph convolution, IAVF generally outperforms the strongest baselines. For example, it achieves 78.53% accuracy on NCI109, corresponding to a relative improvement of 1.03%, and 81.88% on Mutagenicity, corresponding to a relative improvement of 0.69%. These results demonstrate that explicit cross-view interaction enhances graph representations.
Keywords:
Graph convolutional networks (GCNs)
interaction awareness
multiscale fusion
multiview fusion
multiview learning
Journal
IF:
4.9
Papers:
577
Citations:
6.8K