[R-package] [python-package] More convenient API for interaction_constraints

mayer79 commented 3 months ago

Interaction constraints are extremely useful in practice. The API in LightGBM has two aspects that could be improved:

Currently, the Python package allows only integer positions like [[0], [1, 2]]. We could also allow feature names as in XGBoost, or in the R-package of LightGBM.
Features not appearing in any interaction group are silently dropped from the training altogether. Unlike in Scikit-Learn and XGBoost, where missing features can interact (i.e., build an own interaction group). This applies to both Python/R-APIs. Above example could be entered as [[0]], or as [["living_area"]].

Both aspects help to reduce user mistakes and aligns the API with XGBoost (and partly with Scikit-Learn's histogram gradient boosting implementation).

I can make two PRs: one for the R-package, and one for the Python package.

@jameslamb @btrotta

mayer79 commented 3 months ago

@jameslamb could you also add a Python tag? There, we would need two things: translating feature names to int positions, and adding the set of skipped features as additional set.

mayer79 commented 3 months ago

Question to @jameslamb and others: In Python, the interaction constraints are transformed into a string here, just like any other list of list-like object:

https://github.com/microsoft/LightGBM/blob/master/python-package/lightgbm/basic.py,

def _param_dict_to_str(data: Optional[Dict[str, Any]]) -> str:
    """Convert Python dictionary to string, which is passed to C API."""
    if data is None or not data:
        return ""
    pairs = []
    for key, val in data.items():
        if isinstance(val, (list, tuple, set)) or _is_numpy_1d_array(val):
            pairs.append(f"{key}={','.join(map(_to_string, val))}")
        elif isinstance(val, (str, Path, _NUMERIC_TYPES)) or _is_numeric(val):
            pairs.append(f"{key}={val}")
        elif val is not None:
            raise TypeError(f"Unknown type of parameter:{key}, got:{type(val).__name__}")
    return " ".join(pairs)

Is it an option to add a snippet like this?

if key == "interaction_constraints":
  val = _prep_interaction_constraints(vals, feature_names)

_prep_interaction_constraints() would do two things:

Translate feature names to integers (if strings)
Add a list with omitted features

Both would need access to the list of feature_names, so we would need to either pass feature_names = None or the Dataset object to _param_dict_to_str(). Not sure if this the right way to proceed?