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

Abstract

The advancement of deep graph neural networks is hindered by two fundamental challenges in message passing: over-smoothing and over-squashing. Existing graph representation learning methods typically address these problems separately, using homogeneous aggregation strategies that overlook the role-specific nodes in information interaction, making it challenging to preserve critical signals while reducing redundancy. To address these limitations, we propose a novel Diffusion-Perception Co-Learning framework, called DPCoL, which adaptively models the distinct roles of nodes to optimize message passing in deep graph neural networks. DPCoL defines a two-stage message-passing paradigm: (1) Diffusion, where each node dynamically modulates propagation intensity based on feature discriminability, and (2) Perception, where each node selectively identifies task-relevant information from aggregated messages. This co-learning mechanism prevents excessive feature mixing while preserving meaningful signals, ensuring effective long-range message passing without relying on prior assumptions. Unlike previous methods dependent on hyperparameter tuning or explicit smoothing controls, DPCoL employs a learnable diffusion controller and message perceptron, achieving self-adaptive interactions without additional computational overhead. Theoretical analysis validates the convergence properties and expressiveness of DPCoL in addressing the aforementioned problems. Extensive experiments on 13 benchmark datasets, including homophilic, heterophilic, and long-range dependency graphs, demonstrate the robustness and superior performance of the model. DPCoL consistently achieves state-of-the-art results across most scenarios.