FIX: deprecate integer valued numerical features for PDP

[ogrisel]

This is a tentative fix for #30315 and #30378.

The problem is that new pandas versions refuse to assign floating-point values into integer dtyped columns. In PDP, such fractional floating-point values can be generated when creating the grid, even if the original numerical feature is integer valued.

In a first (and later abandoned) version of this PR, I tried to transparently change the dtype of X_eval (or at least the columns where we want to insert those values when X is a dataframe). However, this leads to very pandas and numpy specific code paths and this would need to be expanded for polars later. I have the feeling that this will become unmanageable quickly.

Furthermore, changing the dtype of (some columns of) X_eval means that we are calling the response method of the estimator with different dtypes than the X_train used to fit the estimator. I have the feeling that this can cause weird bugs.

So this PR tries to explicitly deprecate the support of integer values (used as numerical features) in PDPs and later raise a ValueError instead: the error message should be explicit enough to let the user know what to do to update their code. I thought this would be simple, but as you can see in this PR, this is a bit more complex than anticipated.

After creating this PR I am thinking of a final alternative: instead of changing the dtype in X_eval, we could round the fractional

This last alternative will be less intrusive (and in particular will match the implicit behavior of numpy and old pandas versions). But it can be surprising: we generate a fine grid but then it is implicitly coarsified when computing the PDP value. This should be not to complex to implement in a somewhat container agnostic way (using _safe_indexing / _safe_assign and .astype). However, this might be a bit magic. Not sure what is best.

Would love to get feedback from @glemaitre or @lesteve for instance.

[github-actions]

✔️ Linting Passed

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

_{Generated for commit: ca7fb2e. Link to the linter CI: here}

[thomasjpfan]

I'm okay with warning and erroring in the future here.

[glemaitre]

Thanks @ogrisel, I should provide my finding, this could have maybe helped.

I agree with the general idea of the PR. I was leaning towards the same solution and either consider integer dtype as a categorical feature or otherwise, use explicit floating point.

[glemaitre]

I just see that I merged this PR but it targets 1.6 indeed. We could always add it in 1.6.1 because this is kind of a fix + deprecation. @jeremiedbb do you think that we still have time to backport the commit?

[jeremiedbb]

I merged the release PR but I haven't pushed the tag yet so I guess it's doable.