In SAGE, use directly calculated marginal mean as shortcut for empty coalition prediction (or at least featureless learner)

[jemus42]

Currently we actually take a trained learner and feed it a version of the task where every feature has been marginally sampled, which is, in essence, a very inefficient way to calculate mean(target) (in a regression or binary classif case).

Tried to validate one of the shapley assumptions for SAGE (sum(sage_values) ≈ baseline_loss - full_model_loss) and yeah, just using mean(target) lead to a much closer result, which isn't surprising.

[jemus42]

Turns out this is a tad more complicated if we want to over regression/binary/multiclass and arbitrary measures in the current setup.

Gave it a go and decided it's too finicky for no to focus on, I guess SAGE is already complex enough as it is and I don't feel like adding another special case is helping.

[jemus42]

If #54 is done correctly this would also give as quick access to the full and empty coalition losses (or scores at least) computed for each resampling iteration, so depending on the evaluation procedure that either comes in handy or just makes it more complicated.