ENH Allows plotting max class for multiclass in DecisionBoundaryDisplay

[lucyleeow]

Reference Issues/PRs

Towards #27462

What does this implement/fix? Explain your changes.

Allows DecisionBoundaryDisplay to represent all classes for multiclass decision_function and predict_probas by plotting the class with the max response at each point.

Have closely followed the code here: #27291 (comment)

Not 100% on what the colour API should be, open to change.

contour and contourf have a colors parameter, which can be used instead of cmap BUT pcolormesh only has cmap. For simplicity, I've decided to only allow users to pass cmap. This probably makes more sense in this context vs a list of colors to cycle through.

Any other comments?

WIP - need to add tests once we're happy with plot appearance and API

[github-actions]

✔️ Linting Passed

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

_{Generated for commit: 5dc5b7b. Link to the linter CI: here}

[lucyleeow]

I'm also going to try amend example: Logistic Regression 3-class Classifier to show a plot using this enhancement. Though as we use the same X we used to fit the estimator, the plot probably won't look great.
I could instead use Nystroem, though this would only be for the purpose of showing a prettier plot...WDYT @glemaitre ?

[glemaitre]

I would almost argue that we be removing the LogisticRegression plot because it does not have that much values :).

I'm thinking that we could instead modify this one: https://scikit-learn.org/stable/auto_examples/classification/plot_classification_probability.html#sphx-glr-auto-examples-classification-plot-classification-probability-py and add an additional column to the plot with the max. Like this we could remove the OvR LR (it is deprecated), and instead use a Nystroem with LR.

[glemaitre]

Ups I see I didn't post some review that I had. This should be partial comments.

[lucyleeow]

I'm thinking that we could instead modify this one: https://scikit-learn.org/stable/auto_examples/classification/plot_classification_probability.html#sphx-glr-auto-examples-classification-plot-classification-probability-py and add an additional column to the plot with the max. Like this we could remove the OvR LR (it is deprecated), and instead use a Nystroem with LR.

Do you want me to do this in this PR or a separate PR (since removing OvR LR may technically not be considered part of this work)?

[glemaitre]

Do you want me to do this in this PR or a separate PR (since removing OvR LR may technically not be considered part of this work)?

We can do it in this PR and limit the change to this single example.

[glemaitre]

Maybe @ogrisel could have a look since we discussed that feature some time ago.

[lucyleeow]

@ogrisel gentle ping about this 😬 Thank you!

[ogrisel]

@lucyleeow I pushed two commits of changes I wanted to do while reviewing the PR:

• the first selects a qualitative colormap automatically and adjusts it to the observed number of classes;
• the second is a refactoring of the example to show how to read those plots to draw conclusions on the impact of models parameters on the shape of the decision boundary and the magnitude of the predicted probabilities.

I will do a full review tomorrow but this LGTM.

[ogrisel]

This is a pitfall of this method. By default, it would use response_method="decision_function" which is much harder to interpret in my opinion (especially when comparing different model classes).

I think we should change the "auto" policy to use favor predict_proba when available but this should rather be done in a dedicated follow-up PR.

[ogrisel]

BTW, if we use response_method="predict_proba", maybe the vmin=0 and vmax=1 parameters could be set automatically to make the use DecisionBoundaryDisplay terser.

[glemaitre]

It looks indeed like a good suggestion.

[lucyleeow]

FYI @ogrisel I'm just going to fix the merge conflicts

[ogrisel]

Thanks @lucyleeow!

[lucyleeow]

Thanks for finishing this one off @ogrisel !

[glemaitre]

Thanks @ogrisel

[ogrisel]

@lucyleeow If you are looking for follow-up improvements, please consider those two comments:

#29797 (comment)

[lesteve]

FYI this broke the CI when matplotlib is not installed, fix in #30971