Estimator API¶
Public 1.0.0 estimators cover gradient boosting, single decision trees, random forests,
ExtraTrees, CatBoost-like numeric estimators, IsolationForest, and learning-to-rank entries. They
follow the sklearn-style protocol, so normal training, prediction, scoring, and parameter search
can use one method set.
Public names include:
GradientBoostingRegressorGradientBoostingClassifierMPSBoostRegressorMPSBoostClassifierDecisionTreeRegressorDecisionTreeClassifierRandomForestRegressorRandomForestClassifierExtraTreesRegressorExtraTreesClassifierExtraTreeRegressorExtraTreeClassifierCatBoostRegressorCatBoostClassifierIsolationForestMPSIsolationForestLearningToRankRegressor
These estimators follow these sklearn-style entries:
fitpredictpredict_proba, for classifiersscoreget_paramsset_params
Advanced regression objectives are selected through loss:
loss="squared_error"loss="quantile"loss="poisson"loss="tweedie"
Classifiers additionally support probability outputs. Multiclass defaults to the native softmax
path when available; OvR is only an explicit compatibility strategy or staged fallback.
device="auto" chooses CPU or MPS according to environment and workload, and the training summary
records the actual backend so users do not need to guess which path ran.