CALAD: Channel-Aware Contrastive Learning for Anomaly Detection
A new framework called CALAD (Channel-Aware Contrastive Learning) addresses multivariate time series anomaly detection, where labeled data is scarce. Unlike existing unsupervised methods that treat all channels equally, CALAD uses estimated channel relevance to guide contrastive sample construction, focusing on anomaly semantics. Channel relevance is derived from reconstruction errors of a transformer-based autoencoder, identifying channels most influential to anomalies. This enables a channel-wise augmentation strategy for contrastive learning. The paper is available on arXiv under ID 2605.23139.
Key facts
- CALAD is a channel-aware contrastive learning framework for multivariate time series anomaly detection.
- It addresses the problem of scarce labeled data in real-world applications.
- Existing unsupervised approaches often treat all channels equally, diluting anomaly-relevant signals.
- CALAD uses estimated channel relevance to construct contrastive samples.
- Channel relevance is estimated from reconstruction errors of a transformer-based autoencoder.
- The framework distinguishes channels more influential to anomalous behaviors.
- A channel-wise augmentation strategy is designed based on channel relevance.
- The paper is published on arXiv with ID 2605.23139.
Entities
Institutions
- arXiv