Research map / Adversarial robustness and security / Attacks on text and graphs
Alleviating Performance Disparity in Adversarial Spatiotemporal Graph Learning Under Zero-Inflated Distribution
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.
Shows that adversarial training widens the majority/minority performance gap for spatiotemporal graph learning under zero-inflated distributions, and proposes MinGRE, which reweights spatiotemporal gradients with multi-dimensional attention and adds an uncertainty-guided contrastive loss.
Problem
Adversarial training for spatiotemporal graph learning under zero-inflated distributions increases the performance disparity between majority (zero) and minority (non-zero) classes, causing underreporting of critical risk events.
Approach
Combine multi-dimensional (segment, temporal, spatial) attention that reweights spatiotemporal gradients to balance victim-node selection across classes, with an uncertainty-guided supervised contrastive loss that weights minority representations by the negative binomial dispersion parameter.
Key results
- On NYC under STPGD-TNDS, MinGRE reaches Rec-min 34.004 vs 30.974 for AT-TNDS and MAP-min 0.1924 vs 0.1458.
- On the clean NYC dataset, MinGRE reduces Rec-D and MAP-D by 3.6% and 2.3% relative to AT-TNDS.
- Under STPGD-TNDS on NYC, MinGRE lowers Rec-D and MAP-D by 4.0% and 4.2% relative to AT-TNDS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MinGRE instead of AT-TNDS | GSNet · NYC, STPGD-TNDS attack | Rec-min: approximately 9.8% | Helped |
| Use MinGRE instead of AT-TNDS | GSNet · NYC, STPGD-TNDS attack | MAP-min: approximately 31.9% | Helped |
| Apply AT-TNDS rather than natural training | GSNet · NYC | Rec-D: 3.7% | Hurt |
| Apply MinGRE instead of AT-TNDS | GSNet · NYC, clean | Rec-D: 3.6% | Helped |
| Apply MinGRE instead of AT-TNDS | GSNet · NYC, perturbed | MAP-D: 4.2% | Helped |
| Add gradient reweighting (ablation) | AT-TNDS baseline on Chicago · Chicago | performance disparity | Helped |
| Add the spatiotemporal encoder (ablation) | AT-TNDS baseline on Chicago · Chicago | performance | Helped |
| Add uncertainty-guided contrastive loss (ablation) | AT-TNDS baseline on Chicago · Chicago | Rec-D | Helped |
| Add gradient reweighting and uncertainty-guided contrastive loss (ablation) | AT-TNDS baseline on Chicago · Chicago | performance disparity | Helped |
Surprising findings
- Traditional adversarial training (AT-TNDS) increases the majority/minority performance disparity (Rec-D up 3.7%) on NYC, even while it improves average robustness.
- ZID-specific losses (NBL, BMSE) do not consistently achieve the best robust disparity, sometimes underperforming adversarial training on NYC.
Limitations
- Evaluated only on two urban traffic accident datasets (NYC and Chicago) and two target models.
- Several components (victim node selection, attention reweighting) rely on heuristic assumptions and the paper notes residual minority-class difficulty.
Experimental setup
Datasets: NYC traffic accident data Chicago traffic accident data
Benchmarks: NYC Chicago
Metrics: Rec-maj Rec-min MAP-maj MAP-min Rec-D MAP-D
Baselines: AT-Random AT-Degree AT-Pagerank AT-TNDS NT-WRMSE NT-NBL NT-BMSE STPGD-Random STPGD-Degree STPGD-Pagerank STPGD-TNDS
Builds on: STPGD AT-TNDS supervised contrastive loss GSNet Graph WaveNet
Closest papers by meaning
- On the Stability of Graph Convolutional Neural Networks: A Probabilistic Perspective (NeurIPS 2025 · similarity 0.85)
- Refine then Classify: Robust Graph Neural Networks with Reliable Neighborhood Contrastive Refinement (AAAI 2025 · similarity 0.85)
- Value at Adversarial Risk: A Graph Defense Strategy against Cost-Aware Attacks (AAAI 2024 · similarity 0.84)
- GOLD: Graph Out-of-Distribution Detection via Implicit Adversarial Latent Generation (ICLR 2025 · similarity 0.84)
- If You Want to Be Robust, Be Wary of Initialization (NeurIPS 2024 · similarity 0.83)
- Mirage: Model-agnostic Graph Distillation for Graph Classification (ICLR 2024 · similarity 0.83)
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.