DEGNN: Dual Experts Graph Neural Network Handling Both Edge and Node Feature Noise

TL;DR

DEGNN tackles the robustness gap in graph learning by decoupling edge and node-feature processing through two dedicated experts trained with self-supervised contrastive objectives. The Node Feature Expert generates noise-robust embeddings, while the Edge Expert learns a complementary representation to guide edge reconstruction via cosine-similarity-based rewiring, with the downstream GNN leveraging both outputs end-to-end. Across four datasets and multiple noise regimes, DEGNN demonstrates strong robustness to edge and node feature noise, outperforming or competitive with state-of-the-art GSL methods especially on datasets with noisy features. This decoupled, end-to-end framework offers a principled path to robust graph representations in real-world, imperfect graphs, and invites future exploration of more specialized encoders and augmentation strategies for further gains.

Abstract

Graph Neural Networks (GNNs) have achieved notable success in various applications over graph data. However, recent research has revealed that real-world graphs often contain noise, and GNNs are susceptible to noise in the graph. To address this issue, several Graph Structure Learning (GSL) models have been introduced. While GSL models are tailored to enhance robustness against edge noise through edge reconstruction, a significant limitation surfaces: their high reliance on node features. This inherent dependence amplifies their susceptibility to noise within node features. Recognizing this vulnerability, we present DEGNN, a novel GNN model designed to adeptly mitigate noise in both edges and node features. The core idea of DEGNN is to design two separate experts: an edge expert and a node feature expert. These experts utilize self-supervised learning techniques to produce modified edges and node features. Leveraging these modified representations, DEGNN subsequently addresses downstream tasks, ensuring robustness against noise present in both edges and node features of real-world graphs. Notably, the modification process can be trained end-to-end, empowering DEGNN to adjust dynamically and achieves optimal edge and node representations for specific tasks. Comprehensive experiments demonstrate DEGNN's efficacy in managing noise, both in original real-world graphs and in graphs with synthetic noise.

Figures (6)

• Figure 1: A comparison of semi-supervised node classification between GCN and GSL models on Cora and Citeseer dataset when there is noise in the node features.
• Figure 2: The overview of DEGNN and the expert in its pre-training process.
• Figure 3: Accuracy on Citeseer with noise added to either or both of edge and node features.
• Figure 4: Accuracy on Cora with noise added to either or both of edge and node features.
• Figure 5: Comparison of edge homophily ratio and node homophily ratio between noisy graphs and refined graphs by the edge expert on Cora and Citeseer.