Answers are grounded in this paper and cite their sources.
Traffic anomaly detection is critical for urban safety. Existing methods struggle to capture context-dependent normal patterns, often confusing routine fluctuations with true anomalies and causing high false positives. While diffusion models show promise, most treat contextual information as mere auxiliary inputs and fail to achieve componentwise, context-differentiated conditioning. To address these limitations, we propose ConDiST, a context-conditioned spatiotemporal diffusion framework that takes external temporal features and long-term trends as explicit conditioning signals. In the temporal domain, we design a context-conditioned temporal modeling module that uses wavelet decomposition to let each frequency band evolve along a context-dependent path, mitigating cross-scale coupling. In the spatial domain, we develop a context-conditioned spatial modeling module that anchors on fixed geographic priors as a stable basis while learning context-driven structural evolution, enabling dynamic reconfiguration of spatial dependencies. Building on these, we introduce an anomaly scoring method that multiplicatively couples deterministic prediction errors with diffusion-based uncertainty for synergistic enhancement. Extensive experiments on NYC, Chicago, and PeMS benchmarks demonstrate that ConDiST achieves the highest mean AUC on all three datasets, outperforming the strongest baseline by 0.72, 0.88, and 2.21percentage points on NYC, Chicago, and PeMS, respectively.
No takes yet. Share an insight, caveat, or question.
Jiang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: