Alleviating Performance Disparity in Adversarial Spatiotemporal Graph Learning Under Zero-Inflated Distribution

TL;DR

This work tackles robustness and fairness in spatiotemporal graph learning under zero-inflated distributions (ZID), where minority non-zero events are particularly important for safety analyses but underrepresented. It identifies that conventional adversarial training can widen performance gaps between majority and minority classes due to imbalanced gradient flows. To address this, the authors propose MinGRE, a framework combining a Cross-Segment Spatiotemporal Encoder with multi-dimensional gradient reweighting and an uncertainty-guided contrastive loss, yielding balanced minority and majority representations and improved class separability. Empirical results on NYC and Chicago datasets show that MinGRE reduces both natural and robust disparities and outperforms existing adversarial training baselines, supporting more equitable and robust urban risk modeling. The approach integrates a gradient-aware adversarial sample generator and NB-based uncertainty weighting to tailor robustness where risk events are sparse, with potential impact on practical deployment in urban analytics and safety monitoring.

Abstract

Spatiotemporal Graph Learning (SGL) under Zero-Inflated Distribution (ZID) is crucial for urban risk management tasks, including crime prediction and traffic accident profiling. However, SGL models are vulnerable to adversarial attacks, compromising their practical utility. While adversarial training (AT) has been widely used to bolster model robustness, our study finds that traditional AT exacerbates performance disparities between majority and minority classes under ZID, potentially leading to irreparable losses due to underreporting critical risk events. In this paper, we first demonstrate the smaller top-k gradients and lower separability of minority class are key factors contributing to this disparity. To address these issues, we propose MinGRE, a framework for Minority Class Gradients and Representations Enhancement. MinGRE employs a multi-dimensional attention mechanism to reweight spatiotemporal gradients, minimizing the gradient distribution discrepancies across classes. Additionally, we introduce an uncertainty-guided contrastive loss to improve the inter-class separability and intra-class compactness of minority representations with higher uncertainty. Extensive experiments demonstrate that the MinGRE framework not only significantly reduces the performance disparity across classes but also achieves enhanced robustness compared to existing baselines. These findings underscore the potential of our method in fostering the development of more equitable and robust models.

Figures (4)

• Figure 1: Impact analysis of spatiotemporal adversarial training on the ZID dataset NYC. (a) compares recall metrics between natural training and adversarial training approaches. (b) displays the distribution of top-K gradients for both majority and minority classes throughout the adversarial training. Panels (c) and (d) present two-dimensional projections of the learned features for majority and minority classes via AT-TNDS and our proposed method, respectively.
• Figure 2: The overall framework of our proposed MinGRE.
• Figure 3: Effectiveness of sub-modules on NYC datasets.
• Figure 4: Ablation studies of the proposed Adversarial Examples Generation Module and Uncertainty-guided Contrastive Loss Module on Chicago datasets.