I have a dataset of normal points Xtr and I would like to fit a OneClass-SVM for anomaly detection purposes.
from sklearn.svm import OneClassSVM
#fit
ocsvm = OneClassSVM(nu=NU, kernel='rbf', gamma=GAMMA)
ocsvm.fit(X=Xtr)
Then, I would like to assess if every point in the test set Xts is an anomaly:
ocsvm.predict(Xts)
I would like also to determine a degree of anomaly by calculating the distance from the boundary:
ocsvm.decision_function(Xts)
Everything easy so far, but I would like to understand which feature contributes more to causing an anomaly?
How can I determine which of the n variables exceeded the limit of the normality curve defined by the one-class SVM during the training phase?
Should I manually calculate the distance for every x in Xts over all the coordinates (the n features) to understand the actual contribution of each feature?
Do there exist some libraries that already do it?
I have checked sklearn and pyod but it seems that this calculation is not implemented.
