An ensemble detector using multiple random forests on different feature subsets (Random Black Forest)
Citekey | - |
Source Code | own |
Learning type | semi-supervised |
Input dimensionality | multivariate |
An ensemble windowed multi-output forecasting method using random forest regression and random subspace ensembling (requested by RollsRoyce). The forecasting error is used as anomaly score.
The regressor ensebmle is trained on a clean time series to look at a fixed window (train_window_size
points) and predict the next point.
On the test series, the predicted values are compared to the observed ones and the prediction error is returned as anomaly score.
The first train_window_size
points of the test series don't get an anomaly score (are set to NaN
), because no predictions are possible for them.