Graph Neural Networks (GNNs) have become the standard for graph classification tasks, yet they remain remarkably sensitive to structural noise. Even minor perturbations, such as the addition or removal of a few edges, can lead to significant shifts in internal representations and a sharp decline in prediction confidence. This paper introduces MGRL (Multi-View Robust Graph Representation Learning), a framework designed to enhance GNN robustness through multi-granularity contrastive learning and class-view discriminative optimization.
The Robustness Challenge in GNNs
Current GNN architectures face two primary hurdles when dealing with noisy data: Semantic Bias and Confidence Collapse. Semantic bias occurs when a small structural change causes the graph's mathematical embedding to drift away from its original semantic meaning. Confidence collapse refers to the phenomenon where, even if a model correctly labels a perturbed graph, its certainty drops precipitously. For instance, the authors observed that removing just 30% of edges could cause prediction confidence to swing by as much as 60%, rendering the model unreliable for high-stakes applications like bioinformatics.
The MGRL Framework
To address these issues, MGRL employs a dual-view learning strategy that ensures both local stability and global class separation.
1. Instance-View Consistency
The instance-view module focuses on maintaining semantic stability. Rather than relying on manual edge perturbations which might destroy graph topology, MGRL uses neural network dropout to generate multiple "perturbed" views of the same graph. It then applies multi-granularity contrastive learning at both the node and graph levels. To avoid the "false negative trap"—where graphs of the same class are accidentally pushed apart in the latent space—MGRL incorporates an instance weighting mechanism that uses ground-truth labels to refine the contrastive loss.
2. Class-View Discriminative Learning
The class-view module aims to prevent confidence collapse by creating well-defined boundaries between categories. The framework maintains "class prototypes"—representative mathematical centers for each category. By optimizing the distance between graph representations and their corresponding prototypes, the model anchors noisy graphs to stable semantic centers. Furthermore, a "prototype center loss" is introduced to maximize the distance between different class prototypes, ensuring a more compact and discriminative embedding space.
Experimental Evaluation
MGRL was evaluated against state-of-the-art models on eight benchmark datasets from the TUDataset collection, including social networks (IMDB, COLLAB, REDDIT) and bioinformatics (PROTEINS, NCI1).
- Superior Accuracy: MGRL consistently outperformed competitors like GraphCL and G-Mixup. Notably, it achieved a 2.84% accuracy improvement over GraphCL on the COLLAB dataset.
- Visual Robustness: PCA and t-SNE visualizations demonstrate that MGRL maintains highly clustered embeddings even under perturbation. When edges are removed, MGRL's representations remain significantly closer to their class anchors (prototypes) compared to baseline models.
- Confidence Stability: In tests on the Mutagenicity (MUT) dataset, MGRL exhibited the smallest "jitter" in prediction confidence when faced with perturbations, effectively mitigating the confidence collapse issue.
Conclusion
MGRL demonstrates that robustness in GNNs requires more than just average accuracy; it requires a stable and discriminative latent space. By combining instance-level consistency with class-level prototypes, MGRL provides a framework that remains both semantically accurate and confident in the face of structural noise. Future work may explore the integration of domain-specific augmentations to further enhance the model's versatility across diverse graph types.
Summary updated