Reproduce “Harnessing Contrastive Learning and Neural Transformation for Time Series Anomaly Detection” on OpenResearch

About this paper

Katrina Chen, Mingbin Feng, Tony S. Wirjanto

Researchers at the University of Waterloo developed CNT, an unsupervised method for time series anomaly detection that combines a window-based contrastive learning strategy with learnable neural transformations. It achieves improved performance by preventing representation collapse without relying on problematic negative sampling, validated across five industrial datasets.

Key insights

• Learnable neural transformations applied in the latent space effectively prevent representation collapse in contrastive learning for time series, offering a robust alternative to negative sampling.
• A combined objective function, balancing contextual semantic similarity and representation diversity, is crucial for learning meaningful representations of normal time series data.
• Focusing anomaly detection on deviations within a shorter prediction window relative to its historical context allows for the effective identification of subtle, context-dependent anomalies.

Summary by alphaXiv.