Add user friendly string options for interaction constraints in HistGradientBoosting*

[lorentzenchr]

Describe the workflow you want to enable

model_no_interactions = HistGradientBoostingRegressor(
    interaction_cst="no_interactions"
)

model_pairwise_interactions = HistGradientBoostingRegressor(
    interaction_cst="pairwise"
)

instead of

model_no_interactions = HistGradientBoostingRegressor(
    interaction_cst=[[i] for i in range(X_train.shape[1])]
)

model_pairwise_interactions = HistGradientBoostingRegressor(
    interaction_cst=list(itertools.combinations(range(n_features), 2))
)

Describe your proposed solution

• "no_interactions" is straight forward.
• "pairwise" expands to a list that is quadratic in number of features. It might be more memory efficient to use generators internally.

Describe alternatives you've considered, if relevant

No response

Additional context

This was proposed as follow-up in #21020 (comment).

[mzugravu]

Hi,

If no one is taking it I can take this issue. I have to make a contribution as part of one of my lectures at university so I'd be grateful for letting me do it.

However, perhaps I need more explanation. I've just checked the doc of HistGradientBoostingRegressor and there is no such argument as interaction_cst. So you would like to add this parameter and link the key word "no_interactions" or "pairwise" to what you showed in the part instead of. Am I correct? Perhaps you could describe a bit what this interaction means for this case.

Let me know if I got it right and if I can take this issue. (don't hesitate to tell me if this will be too hard for a beginner) Thanks.

[lorentzenchr]

@mzugravu Have a look at the linked PR. Interaction constraints are a new feature that will be released in 1.2. I split this issue into 2 parts. The first part "no_interactions" is the by far easier one.
But if this is your first contribution, I suggest searching for issues with label "good first issue" (or "easy").

[mzugravu]

Ok, thank you. I think I will keep searching it might be too hard for me.

[betatim]

I'll take a look at this

[ogrisel]

And maybe also (possibly in a follow-up PR) using the input feature names of the model when available:

model_no_interactions = HistGradientBoostingRegressor(
    interaction_cst=[
        ("name_of_feature_0", "name_of_feature_42"),
        ("name_of_feature_0", "name_of_feature_1", "name_of_feature_2"),
    ]
)

[lorentzenchr]

@betatim Go ahead.
@ogrisel We should open a new issue for feature names as argument option, for monotonic as well as interaction constraints.

[ogrisel]

@ogrisel We should open a new issue for feature names as argument option, for monotonic as well as interaction constraints.

I agree. Let me do that and I will cross-link back to this issue.