Beyond Soft Masks: Hard-Perturbation Mixup Explainer for Robust GNN Explainability

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences · Depth: Expert, extended

Summary

HPME, a Hard-Perturbation Mixup Explanation framework, enhances Graph Neural Network (GNN) explainability by addressing limitations of existing soft-mask-based methods. These prior approaches often allow label-irrelevant information to persist, leading to out-of-distribution (OOD) problems and reduced explanation fidelity. HPME introduces a generalized Graph Information Bottleneck that uses graph pooling to extract discrete explanatory subgraphs, effectively compressing label-irrelevant components. It also features a novel structural mixup strategy, built on structure-level replacement, to generate in-distribution explanations and mitigate distribution shifts. Extensive experiments on diverse synthetic and real-world datasets demonstrate HPME's state-of-the-art performance, achieving up to a 30.1% improvement in AUC for robust and interpretable explanations across classification and regression tasks.

Key takeaway

For AI Scientists and Machine Learning Engineers developing Graph Neural Network (GNN) applications in high-stakes domains, you should consider HPME to overcome the limitations of traditional soft-mask explainers. Its hard-perturbation and structural mixup strategies provide more robust and interpretable explanations by effectively eliminating label-irrelevant information and mitigating out-of-distribution issues. This approach ensures higher explanation fidelity, crucial for building trust and broader adoption of GNNs in critical applications like drug discovery or fraud detection.

Key insights

HPME uses hard perturbations and structural mixup to generate robust, in-distribution GNN explanations by eliminating irrelevant information.

Principles

• Soft masks in GNN explanations degrade fidelity due to irrelevant information leakage.
• Hard perturbations via graph pooling can enforce explicit information compression.
• Structural mixup mitigates OOD issues by preserving natural graph connectivity.

Method

HPME extends GIB with hard perturbations via graph pooling to extract discrete subgraphs. It then uses a structural mixup strategy, replacing pooled subgraphs between a target and sampled graph, to create in-distribution mixup graphs for prediction loss calculation.

In practice

• Apply graph pooling to extract discrete explanatory subgraphs.
• Use structural replacement for mixup to maintain original graph distribution.
• Evaluate explanation fidelity using AUC-ROC, cosine similarity, and Euclidean distance.

Topics

• Graph Neural Networks
• GNN Explainability
• Information Bottleneck
• Graph Pooling
• Structural Mixup
• Out-of-Distribution Mitigation

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Beyond Soft Masks: Hard-Perturbation Mixup Explainer for Robust GNN Explainability — HPME improves GNN explainability by using hard perturbations and structure-level mixup to eliminate irrelevant information and mitigate OOD issues.
• Understanding Convolutions on Graphs — GNNs process graph-structured data by aggregating and combining node features, enabling powerful predictions on complex systems.
• GJDNet: Robust Graph Neural Networks via Joint Disentangled Learning Against Adversarial Attacks — GJDNet enhances GNN robustness by jointly disentangling node representations and decision spaces against adversarial attacks.
• Community Concealment from Unsupervised Graph Learning-Based Clustering — Community concealment in GNNs depends on boundary connectivity and feature similarity, mitigated by targeted perturbations.
• AIMing for Standardised Explainability Evaluation in GNNs: A Framework and Case Study on Graph Kernel Networks — AIM provides a flexible framework for standardized, comprehensive explainability evaluation across diverse GNN models.
• Graph Rewiring in GNNs to Mitigate Over-Squashing and Over-Smoothing: A Survey — Graph rewiring modifies GNN topology to combat over-squashing and over-smoothing, improving information flow.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.