Problem set & some possible solution

Anomaly detection

Simple time series model AR or ARMA
- predict graph next hour -> error is too big -> anomaly
- estimate model : variance of error
  - distribution assumption, normal -> set threshold based on probability
Time series: LSTM
Isolation Forest, 1 class SVM
AE Neural Net
Supervised
How to deal with seasonal freq
- Seasonal trend decomposition using loess
- Non linear regression to establish threshold around that mid point, based on relationship between the variance and the level
Other statistical method
- change point detection (mean, variance)
  - maximum likelihood estimation
- Statistical Process Control