Improve MinCovDet.fit error when covariance is zero

Ok this is extremely weird, can someone run this code and see if it crashes with that error?

import numpy as np

from sklearn.covariance import MinCovDet

clf = MinCovDet()

data = np.array([0.5, 0.1, 0.1, 0.1, 0.957, 0.1, 0.1,
                 0.1, 0.4285, 0.1]).reshape(-1, 1)
clf.fit(data)

If I change the array to this

data = np.array([0.5, 0.11, 0.1, 0.1, 0.957, 0.1, 0.1, 
                 0.1, 0.4285, 0.1]).reshape(-1, 1)

Then it runs fine

But it seems to crash with any array where there are too many of the same values

This array crashes as well

data = np.array([0.5, 0.3, 0.3, 0.3, 0.957, 0.3, 0.3, 
                 0.3, 0.4285, 0.3]).reshape(-1, 1)

I already checked for NANs and everything, there’s nothing

Using Python 3.6.2 Pandas 0.20.3 Numpy 1.13.1 scikit-learn 0.19.0

Thanks

[albertcthomas]

In this case, the estimated covariance matrix of the support data is equal to 0 and therefore the determinant is equal to 0. We thus have the minimum covariance Determinant and the algorithm should stop as explained in the original paper. The covariance is equal to 0 in your case because the support data are the ones with the same values… If there are fewer ties you have different values in the support data and everything works fine. If you don’t want the robust covariance to be estimated by 0 you may want to increase the support_fraction parameter to increase the number of support data.

[lesteve]

To be honest, just saying “det(cov) = 0, try to increase support_fraction” may be a good enough error message. PR more than welcome!