Dynamic Soft Contrastive Learning for Time Series Anomaly Detection

Abstract

Multivariate time series anomaly detection has been extensively studied in diverse domains. Within the horizon of unsupervised methods, density estimation is regarded as a promising direction. However, the anomalies that distribute closely with normal data in data space impact the desired density estimation results, thereby significantly affecting the discrimination of anomalies. In this work, to tackle this challenge, we propose Dynamic Soft Contrastive Learning (DiSCo). DiSCo generates a representation that better captures the underlying correlations within the data distribution, providing a more compact, informative, and suitable description for the subsequent density estimation task. Furthermore, DiSCo employs a dynamic and learnable distance metric to measure similarities between inputs. The learned task-specific distance metric contributes to minimizing the intra-class distance and maximizing the inter-class distance, thereby assisting the generation of better representation. The efficacy of DiSCo is proved through extensive experiments conducted on four common anomaly detection datasets.