In a typical multivariate scenario, where the time series does not really follow any particular pattern (or distribution), are the techniques like Standardization or MiniMax scaling appropriate ? Or should one rather try making the series Stationary individually, or applying moving average to smooth the series first. My objective is to use LSTM models to be able to predict anomalous behavior in near real-time, so a single mean and a variance for the overall time series, will not be true representative of the data and thus the standard techniques of feature scaling would be tricky. Can anyone share some article or opinion of dealing with such problems ?
How would you define:
"anomalous behavior" - give a scale between 0 and 1
Marcus Hellwig . That would be the ideal situation to have an anomaly defined with a score between [0,1] but I haven't reached there yet. For now, I am defining anomaly based on some threshold. The threshold is itself is set manually by looking at the reconstruction error distribution.
what parameters do you expect f.e.: mean, standarddev., skewness, kurtosis...
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series