Research map / Graph learning / Graph self-supervised learning
Robust Heterogeneous Graph Classification for Molecular Property Prediction with Information Bottleneck
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RHGC, an information-bottleneck-based robust heterogeneous GNN for molecular graph classification that separates label-relevant from label-irrelevant subgraphs at node, semantic, and graph levels, evaluated under structure, feature, and evasion attacks.
Problem
HGNNs for molecular property prediction are highly vulnerable to adversarial perturbations (global direct and local intermediate perturbations), and existing robust methods do not handle heterogeneous graph classification.
Approach
A Node Semantic Purifier uses GSAT-style graph stochastic attention per metapath and HSIC regularization across metapaths, combined with a Global Graph Disentanglement module that uses a Disentangled Graph Information Bottleneck objective (with adversarial training to minimize mutual information between relevant and irrelevant subgraphs) to avoid label-irrelevant information leak without tuning a
Key results
- RHGC improves accuracy by 4.33% on clean data versus the strongest baseline on average, and by 5.06% across the three attack settings.
- Under non-targeted structure and feature attacks, RHGC reaches 82.03% and 83.41% on OGB-molbace versus 78.82% for PGIB.
- Under evasion attack, RHGC achieves 64.67% on PROTEINS, 67.65% on OGB-molbace, and 74.33% on MUTAG, outperforming the best baseline by about 4.89% on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RHGC instead of the strongest reported baseline | graph classification · PROTEINS, OGB-molbace, and MUTAG; clean and non-targeted attacks | accuracy: 81.24±2.11 vs 77.50±2.38; 87.87±1.65 vs 83.61±4.32; 91.52±2.23 vs 86.53±1.97 (clean) | Helped |
| use RHGC instead of the strongest reported baseline | graph classification · PROTEINS, OGB-molbace, and MUTAG evasion attack | accuracy: 64.67±4.27 vs 58.14±4.49; 67.65±3.26 vs 63.53±4.57; 74.33±3.82 vs 70.29±3.68 | Helped |
| add NSP to the GSAT backbone without GGD (ablation) | GSAT backbone · PROTEINS clean | accuracy: 69.64±4.71 -> 74.72±3.54 | Helped |
| add NSP to the GSAT backbone without GGD (ablation) | GSAT backbone · OGB-molbace clean | accuracy: 77.21±3.46 -> 80.23±4.21 | Helped |
| add NSP to the GSAT backbone without GGD (ablation) | GSAT backbone · MUTAG clean | accuracy: 80.88±8.94 -> 84.33±3.71 | Helped |
| add GGD to the GSAT backbone without NSP (ablation) | GSAT backbone · PROTEINS clean | accuracy: 69.64±4.71 -> 78.12±1.96 | Helped |
| add GGD to the GSAT backbone without NSP (ablation) | GSAT backbone · OGB-molbace clean | accuracy: 77.21±3.46 -> 81.61±2.76 | Helped |
| add GGD to the GSAT backbone without NSP (ablation) | GSAT backbone · MUTAG clean | accuracy: 80.88±8.94 -> 86.49±2.43 | Helped |
Limitations
- Only evasion attacks evaluated for targeted scenarios; poisoning attacks not studied due to lack of code.
- Graph reconstructor adapted from GAE; advanced graph reconstruction for heterogeneous graphs left for future work.
- Hyperparameter lambda sensitivity; best at 0.1 chosen on the same datasets.
Experimental setup
Datasets: OGB-molbace MUTAG PROTEINS
Benchmarks: OGB-molbace MUTAG PROTEINS
Metrics: accuracy
Baselines: muxGNN HeGCL NodeSam G-Mixup InfoGraph MGRL GIB VIB-GSL PGIB
Builds on: GSAT disentangled information bottleneck metapath-based HGNN GRABNEL
Closest papers by meaning
- Empowering Active Learning for 3D Molecular Graphs with Geometric Graph Isomorphism (NeurIPS 2024 · similarity 0.87)
- Pre-Training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2025 · similarity 0.87)
- On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.86)
- HAGO-Net: Hierarchical Geometric Message Passing for Molecular Representation Learning (AAAI 2024 · similarity 0.86)
- Position: Graph Learning Will Lose Relevance Due To Poor Benchmarks (ICML 2025 · similarity 0.86)
- Bi-level Contrastive Learning for Knowledge-Enhanced Molecule Representations (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.