Reproduce “Benchmarking Unsupervised Strategies for Anomaly Detection in Multivariate Time Series” on OpenResearch

About this paper

Laura Boggia, Rafael Teixeira de Lima, Bogdan Malaescu

Benchmarking unsupervised anomaly detection methods for multivariate time series, IBM Research Paris-Saclay and Sorbonne Université researchers found the iTransformer architecture, particularly its reconstruction-based variant, emerged as the top performer across 8 out of 10 diverse real-world datasets, leveraging its inverted embedding for superior capture of inter-variate relationships. The study also offers guidelines for anomaly label extraction and addresses the impact of anomalies in training data.

Key insights

• The iTransformer's 'inverted embedding' mechanism, which applies attention across variates, proves highly effective for multivariate time series anomaly detection, particularly for reconstruction-based approaches.
• The most effective anomaly label extraction strategy depends on dataset characteristics, with local 'inclusive OR' suitable for point anomalies and 'majority voting' or 'global' methods preferred for longer anomalous sequences.
• The presence of even a low percentage of anomalies in training data significantly degrades model performance, though alternative loss functions like Huber or Soft-DTW can offer some mitigation.

Summary by alphaXiv.