Flow2GNN: Flexible Two-Way Flow Message Passing for Enhancing GNNs Beyond Homophily.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Message Passing (MP) is fundamental for Graph Neural Networks (GNNs).
• Traditional MP struggles with heterophily graphs due to redundant information.
• Existing heterophily-focused GNNs often have complex designs.

Purpose of the Study:

• To propose a novel, simple message-passing scheme, Flow2GNN, for improved GNN performance on heterophily graphs.
• To enhance GNNs by disentangling and redistributing heterophily information in topology and attribute spaces.
• To demonstrate the flexibility of the proposed method in improving various GNN architectures.

Main Methods:

• Introduced Flow2GNN, a two-way message-passing neural network.
• Developed a disentangled operator to separate in-flow and out-flow topology information.
• Employed an adaptive aggregation model to adjust homophily and heterophily attribute flow.
• Provided theoretical proof that disentangling reduces the generalization gap.

Main Results:

• Flow2GNN outperforms existing state-of-the-art GNNs on heterophily graphs.
• The method significantly improves the performance of GCN, GAT, GCNII, and H2GCN.
• Achieved up to a 25.88% performance increase for GCN on the Wisconsin dataset.
• Demonstrated enhanced GNN capabilities through disentangled message passing.

Conclusions:

• The proposed Flow2GNN offers a simple yet effective approach to handle heterophily in GNNs.
• Disentangling message passing is a viable strategy to improve GNN generalization on heterophilous data.
• Flow2GNN provides a flexible framework that can enhance existing GNN models.