JuRe: Minimal Denoising Network Achieves Top Anomaly Detection Results
I came across a new study on arXiv that introduces JuRe (Just Repair), a simplified denoising network for spotting anomalies in time series data. It uses a single depthwise-separable convolutional residual block with a hidden size of 128 to fix corrupted segments and measures performance with a structural discrepancy function that doesn’t rely on parameters. While it skips attention mechanisms, latent variables, and adversarial features, JuRe impressively secured second place on the TSB-AD multivariate benchmark with an AUC-PR of 0.404 across 180 series from 17 datasets. It also ranked second on the UCR univariate archive with an AUC-PR of 0.198 from 250 series, surpassing all neural baselines in AUC-PR and VUS-PR. The ablation studies underscore the importance of its denoising goal.
Key facts
- JuRe is a minimal denoising network for time series anomaly detection.
- It uses a single depthwise-separable convolutional residual block with hidden dimension 128.
- No attention, latent variables, or adversarial components are used.
- Ranks second on TSB-AD multivariate benchmark with AUC-PR 0.404.
- Ranks second on UCR univariate archive with AUC-PR 0.198.
- Leads all neural baselines on AUC-PR and VUS-PR.
- Training-time corruption is the dominant factor (ΔAUC-PR = 0.047 on removal).
- Published on arXiv with ID 2604.17388.
Entities
Institutions
- arXiv