PERF Slightly speedup MinCovDet.

[anntzer]

support doesn't need to repeatedly converted from a list of indices (from argsort) to a boolean mask (just do it once at the end); furthermore, the distances don't need to be fully sorted (in O(n log n)), rather, only the n_support first indices need to be selected (in O(n)).

Locally, this patch speeds up the following simple benchmark by ~15%.

import sklearn.covariance
import numpy as np
np.random.seed(1)
# unit gaussian plus 10% outliers
t = np.concatenate([np.random.randn(1000, 2), [2, 4] * np.random.randn(100, 2)])
%timeit sklearn.covariance.MinCovDet().fit(t).covariance_

Reference Issues/PRs

What does this implement/fix? Explain your changes.

Any other comments?

[github-actions]

✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: fc772c7. Link to the linter CI: here}

[RahulVadisetty91]

Use np.argsort to ensure exact n_support smallest distances are selected.

support_indices = np.argsort(dist)[:n_support]

[jeremiedbb]

Thanks for the PR @anntzer. Even if it's not a major speed improvement, the changes are minimal so it's a net improvement to me. LGTM

[ogrisel]

The change LGTM. I timed a similar improvement. I expected the improvement to increase when increasing n_samples but it does not seem so. Anyways I agree with @jeremiedbb's comment above.

Please add a changelog entry in doc/whats_new/v1.6.rst.

[anntzer]

Sure, renamed support to support_indices and added a changelog entry.

[ogrisel]

LGTM. Thanks for the follow-up.