Problem: unsupervised anomaly detection
for seasonal KPIs in web applications.
Donut: an unsupervised anomaly detection algorithm based on VAE.
Background:
有的time series data have seasonal patterns occurring at regular intervals.
Data: KPI shapes: seasonal patterns and local variations, noises.
"abnormal": anomalies and missing points; detect missing points is straightforward.
Existing methods suffer from: 这里面简直是胡说八道。
- the hassle麻烦 of algorithm picking
- parameter tuning
- heavy reliance on labels
- unsatisfying performance
- lack of theoretical foundations
Methodology:
VAE is not a sequential model!!!!!!!!!!!!!!!! thus they apply sliding windows.
在训练时，the anomalies and missing points in a testing window x can bring bias to the mapped z, and further make the reconstruction probability inaccurate.
如何避免anomalies and missing points对训练造成的bias:
- missing points. adopt the MCMC-based missing data imputation technique with the trained VAE. 即模拟出missing points的可能值，然后用可能值，代替missing points 的值。
- anomalies
All the algorithms evaluated in this paper compute one anomaly score for each point. A threshold can be chosen to do the decision: if the score for a point is greater than the threshold, an alert should be triggered