I have used DecisionBoundaryDisplay to plot from_estimator of a DecisionTreeClassifier as in this example.

Based on that I would suggest further improvements:

1. add a title to the plot that clearly states which response_method from which estimator was used in the plot
2. add a legend that shows the values (or value regions) of the response method used (percentages for predict_proba, classes for proba, raw values for decision_function)
   • 2b. redefine the representation of the legend that defines the classes, since they look like a scale rather than marking membership

Do you think this would be useful, @glemaitre?

I agree that we can improve the plot. Maybe some of those improvement would be opt-in via parameter. I need to think more about each of them but having proof of concept will help anyway to know if we want it a certain way.