In a Streaming World, Should You Stand Still? A Comprehensive Benchmark of Anomaly Detection in Streams
A large benchmark finds static time-series anomaly detectors beat streaming methods in most realistic settings.
The paper presents a large streaming benchmark of time-series anomaly detection, comparing incremental streaming detectors with static methods after an initial training batch. It also introduces TSB-drift, a real time-series collection meant to isolate cases where distribution drift would justify online updates. Across most settings, static detectors were more accurate than streaming methods, which the authors say were largely adapted from outlier detection and poorly matched to time-series anomalies.
- Compares streaming and static detectors after an initial training batch.
- Introduces TSB-drift, real series built to isolate distribution drift.
- Static methods beat streaming approaches on accuracy in most settings.
- Authors say streaming TSAD designs do not match modern anomaly needs.
From the original article
1 min read at arxiv.orgTime series anomaly detection (TSAD) is increasingly deployed in streaming settings, where data arrive sequentially and may exhibit non-stationarity. As a result, several works from the recent literature propose streaming anomaly detection methods that rely on incremental updates to adapt over time.
By Magali Parrino · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.