I have a dataset with 100+ features, upon which I test GMM to detect anomalies. For example, I add some Gaussian noise to 5-6 features of 100 points. GMM detects the points easily, but the next suggested step is to develop an algorithm to locate the features with noise. This is where I got stuck.
Outlier score returned by the sklearn is calculated as a sum for all the dimensions of a datapoint. I tried to retrieve internal variables to understand the process of the Gaussian log-likelihood calculation, which underlies the outlier score and somehow segregate features which have outstanding values, but that was not successful. I suspect this has something to do with the way covariance matrices are calculated.
I would be happy to get some hints on where to look at either inside the GMM algorithm or suggestions on some post-detection analysis methods.
