A diffusion-perception co-learning framework for deep graph neural networks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Deep graph neural networks face challenges like over-smoothing and over-squashing in message passing.
• Current methods often address these issues separately, using homogeneous aggregation that neglects node-specific roles and signal preservation.

Purpose of the Study:

• To propose a novel Diffusion-Perception Co-Learning (DPCoL) framework to optimize message passing in deep graph neural networks.
• To adaptively model distinct node roles for improved information interaction and signal preservation.

Main Methods:

• DPCoL employs a two-stage message-passing paradigm: Diffusion (nodes modulate propagation based on feature discriminability) and Perception (nodes selectively identify task-relevant information).
• A learnable diffusion controller and message perceptron enable self-adaptive interactions without additional computational overhead.
• Theoretical analysis confirms DPCoL's convergence and expressiveness in addressing over-smoothing and over-squashing.

Main Results:

• DPCoL effectively prevents excessive feature mixing while preserving meaningful signals for long-range message passing.
• Experiments on 13 diverse datasets (homophilic, heterophilic, long-range) demonstrate DPCoL's robustness and superior performance.
• The model consistently achieves state-of-the-art results across various graph scenarios.

Conclusions:

• DPCoL offers a novel, adaptive approach to message passing in deep graph neural networks, overcoming limitations of existing methods.
• The framework achieves superior performance and robustness without relying on hyperparameter tuning or explicit smoothing controls.