Xi'an Jiaotong's HPME Fixes the Flaw Hiding in Every GNN Explainer
HPME replaces soft edge masks with hard graph pooling and structure-level mixup, achieving up to 30% AUC gains over prior GNN explainers across 13 benchmarks.
Takeaways
- HPME is a new post-hoc GNN explainer that replaces soft edge masks with hard graph pooling, eliminating residual label-irrelevant information that leaks into prior methods.
- A novel structural mixup strategy swaps pooled subgraphs between graphs at the same node positions, avoiding the synthetic edges that cause distribution shift in MixupExplainer and ProxyExplainer.
- The framework is grounded in a generalized Graph Information Bottleneck with provable information-capacity bounds from the pooling operation.
- HPME achieves up to 30.1% AUC improvement over the best baseline, averaging 11.4% on classification and 13.4% on regression across 13 datasets.
- Evaluated on 13 benchmarks (7 classification, 6 regression) including molecular datasets like Benzene and Crippen, against 9 baselines including GNNExplainer, PGExplainer, and ProxyExplainer.
- Code is available; the method wraps any pre-trained GNN without retraining and is implemented in PyTorch Geometric.
Graph Neural Networks are increasingly deployed in high-stakes settings like drug discovery, fraud detection, and healthcare. But getting a GNN to tell you why it made a prediction is still an unsolved problem. A new paper from Xi'an Jiaotong University and collaborators takes a hard look at a subtle flaw shared by nearly every modern GNN explainer, and proposes a cleaner fix.
The work introduces HPME (Hard-Perturbation Mixup Explainer), a post-hoc explanation framework that replaces the standard "soft mask" approach with discrete, pooling-based subgraph extraction and a novel structure-level mixup strategy. The result is state-of-the-art explanation accuracy across 13 datasets spanning both graph classification and regression tasks.
The soft mask trap
Most GNN explainers work by learning a mask over the edges of a graph. The mask assigns each edge a weight between 0 and 1, and the edges with the highest weights are presented as the explanation. This is called a soft mask. The problem is that soft masks never truly zero out irrelevant edges. Even after optimization, small residual weights remain on edges that have nothing to do with the prediction.
These approaches typically rely on soft masks, which are inherently unable to fully eliminate label-irrelevant information, allowing redundant structures to leak into the mixup process and hindering the resolution of the OOD problem, thereby degrading explanation fidelity.
A second, related problem is the out-of-distribution (OOD) issue. When you extract a small explanatory subgraph from a full graph and feed it back into the GNN for evaluation, the model was trained on full graphs, not sparse subgraphs. The subgraph sits in a completely different part of the input space, so the GNN's predictions on it are unreliable. Prior methods tried to fix this with a mixup strategy: blend the explanatory subgraph back with a "label-irrelevant" complement to reconstruct something that looks like a full graph. But because soft masks leak residual information, the complement isn't truly label-free, and the resulting mixup graph still drifts away from the original distribution.