Anomalies in Multivariate Time Series Benchmarks Are Mostly Univariate

TL;DR

Multivariate time series data, which involves multiple variables tracked over time, often contains anomalies that are primarily driven by single-variable issues. Researchers analyzed existing benchmarks and discovered that most anomalies could be attributed to univariate (single-variable) factors rather than multivariate interactions.

Key Takeaways

• Evaluate your models against univariate anomaly detection benchmarks.
• Consider developing datasets with diverse anomaly structures.
• Test channel-dependent models for better anomaly detection.
• Revisit assumptions about cross-channel correlations in your analyses.