Jump-Starting Multivariate Time Series Anomaly Detection for Online Service Systems
summary
JumpStarter applies Compressed Sensing with shape-based clustering and outlier-resistant sampling to multivariate time series, reconstructing clustered metric blocks and scoring anomalies by the reconstruction error with an EVT-based threshold. It reports an average F1 of 94.12% and a claimed initialization time of about 20 minutes on datasets from two Internet companies, spots network-induced long response times and software-change hops, runs online with a sliding window, and the code plus datasets are open sourced.
AI-written summary · about
chapters (20)
- 0:00Intro
- 0:31Service Reliability is Important
- 0:53Real-World Revenue Loss
- 1:08Univariate Time Series (UTS)
- 1:58Multivariate Time Series (MTS)
- 3:06Deep Learning Based Approaches: Long Initialization Time
- 3:40Incremental Retraining Offline
- 4:05Incremental Retraining Cannot Ensure Satisfactory Performance
- 4:38Outline
- 5:29Two Strawman Solutions Using CS
- 6:02Problem of Random Gaussian Sampling
- 7:02JumpStarter Overview
- 7:38JumpStarter Offline Processing
- 9:00Outlier-Resistant Sampling
- 10:20JumpStarter Online Processing
- 10:44Compressed Sensing Reconstruction
- 11:08JumpStarter Initialization Time: 20 mins
- 11:59Evaluation: Accuracy
- 12:15Evaluation: Efficiency
- 12:54Conclusion
related talks
45mAnomaly Detection in Time Series
01h 29msktime: time series anomaly detection, changepoint detection, segmentation
28mUnsupervised Real-Time Stream-Based Novelty Detection Technique
52mDevil in the Haystack
19mAutomatic Metric Screening for Service Diagnosis
36mDrift Detection on Irregular Time Series with Multiple Non-Uniform Seasonal Patterns
21mCross-dataset Time Series Anomaly Detection for Cloud Systems
20mSmart Monitoring System for Anomaly Detection on Business Trends