Make it easy to use different color for mean in plot_partial_dependence

It would be great to be able to give a different color or scheme to the average in pd plots. For example, this plot is actually quite interesting in terms of ICE showing trends that are not in the average: (yes this is real data and I’m excited that ICE is useful!) I know which one is the average, but it’s kinda hard to see. If I set subsample to the default, it’s certainly impossible to see. I tried to get the line out of the lines_ but I’m struggling a bit (and I think something is off there EDIT: I was on a pre #18359 commit).

Using the example from the docs you can do:

from sklearn.ensemble import RandomForestRegressor
from sklearn.datasets import fetch_california_housing
from sklearn.inspection import plot_partial_dependence

X, y = fetch_california_housing(return_X_y=True, as_frame=True)
features = ['MedInc', 'AveOccup', 'HouseAge', 'AveRooms', 'AveOccup']
est = RandomForestRegressor(n_estimators=10)
est.fit(X, y)
display = plot_partial_dependence(
       est, X, features, kind="both", subsample=100,
       n_jobs=3, grid_resolution=20, random_state=0
)
display.plot(n_cols=5)
for i in range(display.lines_.shape[1]):
    display.lines_[0,i,-1].set_color('k')
    display.axes_[0, i].legend()  # needed to update the legend color

I don’t think the order of the lines in lines_ is documented.

[glemaitre]

We need to find the right API because we already have line_kw. I don’t know if it would make sense to deprecate it and add ice_lines_kw and pd_line_kw?

[NicolasHug]

or not deprecate but let ice_lines_kw and pd_line_kw take priority over line_kw (i.e. override)? This way one can still control all lines with line_kw while having more fine-grain control with the new params