FIX Fix adjusted_mutual_info_score numerical issue

[glevv]

Reference Issues/PRs

fixes #30950

What does this implement/fix? Explain your changes.

• Make clipping on the values inside adjusted_mutual_info_score symmetric, i.e. both denominator and nominator are now clipped;
• Add shortcut to calculations, to return 0 when there is no split in the data;
• Add non-regression test.

Any other comments?

I'm not sure this is the best way forward

[github-actions]

✔️ Linting Passed

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

_{Generated for commit: 0b40829. Link to the linter CI: here}

[lesteve]

Nice, thanks a lot for the investigation and the fix!

[glevv]

Is it possible to change this construction

if denominator < 0:
    denominator = min(denominator, -np.finfo("float64").eps)
else:
    denominator = max(denominator, np.finfo("float64").eps)

to
denominator = max(normalizer - emi, np.finfo("float64").eps)

Since normalizer should always be >= than emi?

[lesteve]

LGTM, thanks for the fix!

[lesteve]

@virchan if you have some spare bandwitdth and you want to have a second look at this PR, this would be super useful 😉. This seems to fix a real bug in a not too controversial manner.

[virchan]

Thank you for the PR, @glevv.

The proposed fix makes sense to me. I just have a few minor nitpicks:

[glevv]

I had to change some tests to incorporate the changes, so I'm re-requesting review

[virchan]

LGTM! Thanks @glevv!

[lesteve]

Thanks @glevv for the fix and @virchan for the review 🙏!

I tweaked the changelog a bit and enabled auto-merge so this PR will be merged when CI is green!