Contrast to Detect: Dynamic Graph Contrastive Regularization for Unsupervised Anomaly Detection in Multivariate Time Series
Two new research papers propose novel unsupervised learning frameworks for detecting anomalies in multivariate time series data. The first, CALAD, introduces a channel-aware contrastive learning approach that prioritizes anomaly-relevant channels to improve signal detection. The second, ContrastAD, utilizes dynamic graph contrastive regularization, adapting to evolving inter-variable dependencies and using structural evolution as a learning signal. Both methods demonstrate superior performance on various real-world datasets compared to existing techniques, particularly in scenarios with scarce labeled data and distribution shifts. AI
IMPACT These novel unsupervised learning frameworks offer improved accuracy for anomaly detection in critical systems where labeled data is scarce.