LGTM, thanks!

Sorry this took me so long to get to!
This is a much nicer graph, thank you!

I couldn't comment on the line but maybe we could add the name of the category handling strategy in the bullet list at the top, i.e.:

- "Dropped": dropping the categorical features- "One Hot": using a :class:`~preprocessing.OneHotEncoder`- "Ordinal": using an :class:`~preprocessing.OrdinalEncoder` and treat categories as  ordered, equidistant quantities- "Native": using an :class:`~preprocessing.OrdinalEncoder` and rely on the :ref:`native  category support <categorical_support_gbdt>` of the  :class:`~ensemble.HistGradientBoostingRegressor` estimator.

Do you think it's worth mentioning somewhere on the graph that the error bars are 1 standard deviation or is it obvious what the error bars are?

I saw thecomment on addingTargetEncoder but maybe that could be left to another PR...?

It's nice that we make it clear to the user where the best models are (with the arrow and 'best models' text), but it does confuse me at first:

• I expect it to be pointing to something, e.g., a scatter point. We could add a circle that it points to, but this is not necessarily clear either
• The graph axes does not start at 0,0 so it's possibly mis-leading?

I can't think how to improve this. Maybe it is clear enough if people read the label titles, that smaller error/faster is better? Maybe we could add to the x and y axes labels, e.g., an arrow; "<- faster fitting" and "↓ better model/lower error"). Or just explaining it in the text may be enough?

Edit: Maybe we could just have the text 'Best models' in the bottom left corner and no arrow ?

LGTM thanks!

Sorry I meant the words 'faster fitting'/'lower error' may be added to the x and y axis labels.
They do also make sense here, but maybe they are too long inside the graph and 'Best models' was better?