✔️ Linting Passed

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

_{Generated for commit: 69f33db. Link to the linter CI: here}

cc @ogrisel since he had the original comment.

In #31062 (comment) it was suggested to add TargetEncoder to the benchmark, but I realized there's already an example comparing such strategy in the scenario of high cardinality, where it is the most useful.

I think it's worth adding it even to this example. I would expect this method to be among the best even when the cardinality of the categorical features is not that large.

Maybe we can also rework the above mentioned TargetEncoder example?

How would you rework it?

Even merge both examples?

That sounds challenging to do. The examples on wines dataset is already quite slow to run so expanding it and using cv=5 instead of cv=3 would make it even slower. At the same time, it's good to have a categorical feature engineering example that runs on real data with a mix of both high and low cardinality features.

The underfitting analysis of this faster example is nice and complementary to the other. I think I would keep both.

I think it's worth adding it even to this example. I would expect this method to be among the best even when the cardinality of the categorical features is not that large.

@ogrisel these are the results for both plots when adding the TargetEncoder. They seem to be too noisy with the default TargetEncoder(cv=5):

How would you rework it?

To keep both examples, I would cross-reference each other, use the same plotting function and possibly add the native support to the benchmark.

@ArturoAmorQ can I merge?

@lorentzenchr sure, thanks!

@ogrisel these are the results for both plots when adding the TargetEncoder. They seem to be too noisy with the default TargetEncoder(cv=5):

@ArturoAmorQ I don't think the size of the error bars of the TargetEncoder pipeline are significantly larger than the other alternatives. I still think it would be interesting to feature TargetEncoder encoder in this example, as it's usual a competitive alternative to the others (in terms of Pareto optimality) and can furthermore naturally handle both low and high cardinality categories. I think the examples are a useful way to make good alternative discoverable and explain the pros and cons.