TL;DR
This paper presents the first systematic empirical study on how label noise impacts post-hoc Graph Neural Network (GNN) explainers. The authors reveal a troubling disconnect: even minor label noise that has zero impact on a GNN's classification accuracy can severely degrade the quality of its explanations, though a unique "U-shaped" recovery occurs at extreme noise levels.
What Problem Does This Paper Address?
Graph Neural Networks (GNNs) are highly effective for graph-structured data but operate as black boxes. To trust their predictions in critical domains like bio-medicine or social network analysis, researchers deploy post-hoc explainers (e.g., GNNExplainer) to highlight the specific nodes, edges, or subgraphs that drove the model's decision.
While substantial research has focused on making GNN predictions robust to adversarial attacks and noisy data, the robustness of the explainers themselves has been entirely overlooked. This paper tackles a fundamental question: If the labels used to train a GNN are noisy, can we still trust the post-hoc explanations generated from that model?
Methods
The authors build a unified evaluation framework to test two benchmark post-hoc explainers (GNNExplainer and PGExplainer) across two widely used GNN architectures (GCN and GIN).
To simulate real-world data imperfections, they introduce varying levels of label noise ($\lambda$) into the training data by randomly flipping graph labels. They then evaluate the quality of the generated explanations using two standard, label-agnostic metrics:
- Fidelity+ ($Fid^+$): Measures how much the GNN's prediction confidence drops when the identified "important" features are removed. Higher is better.
- Fidelity- ($Fid^-$): Measures how well the GNN maintains its prediction confidence when only the important features are retained. Lower is better.
Datasets
The framework was evaluated on four graph classification datasets:
- MUTAG (Real-world): 4,337 molecule graphs labeled according to their mutagenic effect.
- Graph-Twitter (Real-world): 6,940 text-derived sentiment graphs.
- BA-2motifs (Synthetic): 1,000 Barabasi-Albert (BA) graphs attached to either a "House" or "Cycle" structural motif.
- BA-Multi-Shapes (Synthetic): 1,000 BA graphs attached to "House", "Wheel", or "Grid" motifs.
Key Insights and Ideas
- GNN Robustness Does Not Guarantee Explainer Robustness: At low noise levels ($\lambda \le 20\%$), GNN classification accuracy remains perfectly stable. However, the internal predicted probabilities shift subtly. While these shifts aren't large enough to change the final classification label, they heavily corrupt the inputs to the post-hoc explainers, causing explanation fidelity to plummet.
- The U-Shaped Recovery Phenomenon: As label noise scales past 50%, explanation quality surprisingly begins to recover. At 100% noise (where every single label is inverted), the explainer performs just as well as it does with 0% noise. This happens because the GNN learns the exact same structural motifs, it simply maps them to the opposite inverted label.
- Fidelity- ($Fid^-$) is a Broken Metric for Robustness: The authors point out a critical flaw in standard evaluation. When labels are highly ambiguous (around 50% noise), GNN predictions become uncertain and hover near zero. This mathematically forces $Fid^-$ to approach zero, falsely indicating high-quality explanations when the explainer is actually failing completely.
Key Results
- High Sensitivity to Minor Noise: Across almost all datasets, introducing just 20% label noise caused massive drops in $Fid^+$, even when GNN accuracy didn't budge.
- Complete Failure at 50% Noise: At the 50% noise threshold, explainers completely fail to identify ground-truth structural motifs, such as the mutagenic $NO_2$ chemical group in the MUTAG dataset.
Limitations and Future Work
As a preliminary exploration, this study is limited to standard, vanilla post-hoc explainers and basic graph classification tasks. The authors note several vital directions for the community moving forward:
- Designing Noise-Robust Explainers: There is an urgent need for new post-hoc explanation algorithms explicitly engineered to withstand training label perturbations.
- Rethinking Evaluation Metrics: Because $Fid^-$ fails under high noise, researchers must develop more comprehensive and reliable metrics to evaluate explanation quality.
- Scaling Up: Testing these dynamics on larger-scale benchmark datasets and node-level classification tasks remains an open area of research.
Summary updated