Univariat Outlier Probabilities and thresh automatisation for flagUniLOF
Adds thresh-parameter dependent dispatch to uni variate probabilistic local outlier factor (Local Outlier Probability (LOP)) to flagUniLOF:
This mainly serves the issue of making thresh more interpretable and also redundant (optional) parameter
Standard calling now appropriate with just parameter n
flagUniLOF(field, n=N)Context
LOP derives Outlier Probabilities between 0 and 1 from the outlier factor distributions and thus allows more intuitive threshing. (Also LOP seems to result in a globally more valid ranking, as presented in [1])
-> when assigning thresh=None (default), flagUniLOF switches to local outlier probabilities mode and estimates cut-off probabilities with a k-means cluster-separation approach
-> when assigning thresh in (0,1), flagUniLOF switches to local outlier probabilities mode and calculates probabilities to be cut off, instead of scores
-> when assigning thresh in (-1,0), flagUniLOF switches to local outlier probabilities mode with corruption cap, assuming that not more than thresh percentage of the data are anomalous and therefrom estimates cut-off probability
-> when assigning thresh in (-inf,-1), flagUniLOF switches to local outlier probabilities mode with corruption cap, assuming that not more than thresh number of saples the data are anomalous and therefrom estimates cut-off probability