TRed-GNN: A Robust Graph Neural Network with Task-Relevant Edge Disentanglement and Reverse Process Mechanism

Graph Neural Networks (GNNs) capture complex information in graph-structured data by integrating node features with iterative updates of graph topology. However, they inherently rely on the homophily assumption—that nodes of the same class tend to form edges. In contrast, real-world networks often exhibit heterophilous structures, where edges are frequently formed between nodes of different classes. Consequently, conventional GNNs, which apply uniform smoothing over all nodes, may inadvertently aggregate both task-relevant and task-irrelevant information, leading to suboptimal performance on heterophilous graphs. In this work, we propose TRed-GNN, a novel end-to-end GNN architecture designed to enhance both the performance and robustness of node classification on heterophilous graphs. The proposed approach decomposes the original graph into a task-relevant subgraph and a task-irrelevant subgraph and employs a dual-channel mechanism to independently aggregate features from each topology. To mitigate the interference of task-irrelevant information, we introduce a reverse process mechanism that, without compromising the main task, extracts potentially useful information from the task-irrelevant subgraph while filtering out noise, thereby improving generalization and resilience to perturbations. Theoretical analysis and extensive experiments on multiple real-world datasets demonstrate that TRed-GNN not only achieves superior classification performance compared to existing methods on most benchmarks, but also exhibits strong adaptability and stability under graph structural perturbations and over-smoothing scenarios.