Exploiting Neighbor Effect: Conv-Agnostic GNN Framework for Graphs With Heterophily

Area of Science:

• Machine Learning
• Graph Neural Networks
• Network Science

Background:

• Graph neural networks (GNNs) typically excel on homophilic graphs but struggle with heterophilic graphs due to the homophily assumption.
• Existing metrics inadequately explain GNN performance on heterophilic datasets, suggesting not all inter-class edges are detrimental.
• The role of inter-class edges and neighbor information in GNNs requires deeper investigation for heterophilic graph learning.

Purpose of the Study:

• To re-evaluate the heterophily problem in GNNs using a novel metric based on von Neumann entropy.
• To investigate the impact of inter-class edge feature aggregation from a comprehensive neighbor perspective.
• To propose a framework enhancing GNN performance on heterophilic graphs by learning node-specific neighbor effects.

Main Methods:

• Introduced a new metric using von Neumann entropy to analyze GNN behavior on heterophilic graphs.
• Developed the Conv-Agnostic GNN framework (CAGNNs) that decouples node features into discriminative and aggregation components.
• Implemented a shared mixer module within CAGNNs to adaptively learn and incorporate neighbor effects for each node.

Main Results:

• CAGNNs significantly improve GNN performance on nine benchmark datasets, particularly heterophilic ones.
• Achieved average performance gains of 9.81% (GIN), 25.81% (GAT), and 20.61% (GCN) over baseline models.
• Extensive ablation studies and robustness analyses confirmed the framework's effectiveness, robustness, and interpretability.

Conclusions:

• The proposed metric and CAGNN framework offer a more nuanced understanding and effective solution for GNNs on heterophilic graphs.
• CAGNNs act as a versatile plug-in component, enhancing the performance of various GNN architectures.
• The findings challenge the universal limitations of GNNs on heterophilic data and highlight the importance of adaptive neighbor aggregation.