CoAD: Unifying Classification and Reconstruction for Time Series Anomaly Detection
A new framework called CoAD, proposed in arXiv paper 2605.26193, combines Outlier Exposure (OE) and Masked Autoencoder (MAE) paradigms for time series anomaly detection. OE-based methods suffer from poor generalization, while MAE-based methods face masking misalignment. CoAD uses a classification module to generate probability-informed soft masks for the reconstruction module, addressing both limitations. The framework leverages complementary strengths of classification and reconstruction, improving detection of subtle and prolonged anomalies. The paper challenges the effectiveness of popular deep learning methods for TSAD.
Key facts
- arXiv paper 2605.26193 proposes CoAD framework
- CoAD unifies Outlier Exposure and Masked Autoencoder paradigms
- OE-based methods have poor generalization
- MAE-based methods have masking misalignment issues
- Classification module generates soft masks for reconstruction
- Aims to detect subtle and prolonged anomalies
- Challenges effectiveness of popular deep learning methods for TSAD
- Published on arXiv
Entities
Institutions
- arXiv