[ENH] add CatBoostCvExperiment integration

[SimonBlanke]

The CatBoost estimator seems different enough from sklearn, that we should add a separate integration for it into Hyperactive. The CatBoostCvExperiment could look similar to this:

class CatBoostCvExperiment(BaseExperiment):
    _tags = {
        "python_dependencies": "catboost",
        "property:higher_or_lower_is_better": "lower",
    }
    def __init__(
        self,
        pool,
        iterations=100,
        fold_count=5,
        early_stopping_rounds=None,
        partition_random_seed=0,
        type="Classical",
    ):

[coder-jayp]

I'd like to work on this

My current plan:

• Add a CatBoostCvExperiment class inheriting from BaseExperiment
• Require a pool argument, along with the other CatBoost CV parameters mentioned above
• Use catboost.cv() internally to run the cross-validation instead of the sklearn-style .fit() flow

I’m aiming to keep the initial implementation minimal and aligned with the existing experiment interfaces, and then iterate based on feedback.

Please let me know if this approach looks reasonable, or if there’s an existing experiment class you’d recommend.

[digvijay-y]

@SimonBlanke Hi, I would like to contribute to this project, I will keep an eye on the issue section for any new issues, I'm open to take incomplete work or PRs, it would be a great help for me to get much familiar with the codebase for future contributions.

[passionworkeer]

A practical shape for this integration is to treat catboost.cv() as the execution backend, but normalize its output into the same experiment result contract as the sklearn-style experiments.

Two details seem important:

1. accept Pool directly so categorical/text feature metadata survives intact,
2. resolve the optimization target from the final row of est--mean (and possibly keep est--std alongside it for inspection/debugging).

I’d also consider passing CatBoost params as a dict/kwargs payload instead of hardcoding too many constructor args up front. That keeps the first version small while still supporting objectives/metrics that differ between classification and regression.