Heterophilous distribution propagation for Graph Neural Networks.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Graph Neural Networks (GNNs) excel at representation learning by aggregating neighborhood information, relying on the homophily assumption.
• Real-world graphs often violate homophily, necessitating specialized methods like heterophilous graph neural networks (HeterGNNs).
• Existing HeterGNNs face challenges in effectively partitioning neighborhoods and modeling heterophily.

Purpose of the Study:

• To propose a novel approach, Heterophilous Distribution Propagation (HDP), to address limitations in current HeterGNNs.
• To enhance GNN performance on graphs where nodes exhibit dissimilar behaviors despite proximity.
• To improve representation learning in complex, real-world graph structures.

Main Methods:

• HDP adaptively partitions neighbors into homophilous and heterophilous components using pseudo-assignments during training.
• A trusted prototype contrastive learning paradigm with an orthogonality-oriented constraint is employed to learn heterophilous neighborhood distributions.
• A novel semantic-aware message-passing mechanism propagates both homophilous and heterophilous patterns.

Main Results:

• Extensive experiments were conducted on 9 benchmark datasets with varying homophily levels.
• HDP demonstrated superior performance compared to representative baseline methods on heterophilous datasets.
• The proposed method effectively handles graphs where the homophily assumption is violated.

Conclusions:

• Heterophilous Distribution Propagation (HDP) offers a significant advancement for Graph Neural Networks in heterophilous settings.
• The adaptive neighborhood partitioning and semantic-aware message passing are key to HDP's effectiveness.
• HDP provides a robust framework for representation learning on complex, real-world graphs.