Overlapping inference windows boost time series anomaly detection

Researchers have investigated the impact of inference windowing techniques on the performance of reconstruction-based anomaly detection in time series data. Their study, conducted on the TSB-AD benchmark and the UCR archive, found that using overlapping inference windows consistently improves anomaly detection performance across various models, including PCA, AutoEncoder, TimesNet, and Transformer variants. The improvements averaged up to 28% and altered the comparative rankings of different methods, highlighting the critical role of inference choices alongside model architecture and training. AI

IMPACT Standardizes evaluation protocols, potentially improving comparability and reproducibility of anomaly detection research.

RANK_REASON Academic paper presenting novel methodology and benchmark results for time series anomaly detection. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• AutoEncoder
• Guillaume Coulaud
• PCA
• TimesNet
• Transformer
• TSB-AD
• UCR archive

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →