TBoostClassifier¶

class TBoostClassifier(n_trees=4000,
                       learning_rate=0.05,
                       lambda_=1.0,
                       *,
                       lambda_scale_invariant=False,
                       l1_leaf=0.0,
                       min_split_gain=0.0,
                       max_delta_step=None,
                       max_delta_step_gated=None,
                       max_bin=254,
                       objective="logistic",
                       tweedie_rho=1.5,
                       min_data_in_leaf=None,
                       min_sum_hessian_in_leaf=0.0,
                       min_weight_sum_in_leaf=0.0,
                       path_smooth=None,
                       subsample=None,
                       colsample_bytree=0.8,
                       learning_rate_decay=0.0,
                       validation_fraction=0.1,
                       early_stopping_rounds=500,
                       early_stopping_adaptive=1.5,
                       early_stopping_min_delta=0.0001,
                       interaction_gain_hurdle=2.0,
                       interaction_gain_hurdle_mode="adaptive",
                       graduate=None,
                       graduation_alpha=None,
                       graduation_high_order_alpha=0.0,
                       leaf_refine_steps=4,
                       leaf_refine_backtracks=4,
                       refine_closed_form_tier2=True,
                       incremental_mu=False,
                       mvs_min_rows=1,
                       hist_precision=None,
                       n_bags=8,
                       bag_subsample=0.8,
                       cell_refit_base=None,
                       cell_refit_gamma=2.0,
                       ridge_refit_l2=None,
                       ridge_refit_max_iter=5,
                       nesterov=False,
                       dart_drop_rate=None,
                       random_strength=0.0,
                       reanchor=None,
                       reanchor_slope=None,
                       max_interaction_order=3,
                       max_depth=3,
                       table_budget_cells=None,
                       table_budget_order_shrink=2.0,
                       seed=0,
                       n_jobs=None,
                       monotone_constraints=None,
                       categorical_features=None,
                       cat_smooth=None,
                       cat_target=None,
                       cat_leakage=None,
                       cat_n_perms=1,
                       cat_k=5,
                       cat_min_data_per_group=10.0,
                       cat_direct_max_levels=16,
                       cat_channels=None,
                       cat_count_min_levels=20,
                       cat_class_freq_min_levels=3,
                       prune=True,
                       prune_validation_fraction=0.15,
                       prune_se_rule=0.0,
                       prune_n_folds=5,
                       prune_refit_full=False,
                       prune_rebalance=True,
                       ref_measure=None,
                       measure_floor=0.001,
                       prune_guard=True,
                       prune_guard_tol=0.05,
                       prune_drop_z=2.0,
                       prune_keep_budget=32,
                       prune_fold_fidelity=False,
                       prune_guard_z=0.0,
                       prune_guard_z_dn=2.0,
                       prune_guard_tol_floor=0.005,
                       multiclass_prune_guard=True,
                       multiclass_prune_sel_bags=1,
                       multiclass_prune_cv=True,
                       multiclass_prune_guard_floor=0.002,
                       prune_box_budget=0,
                       prune_lambda_boxes=0.0,
                       prune_table_budget=0,
                       prune_lambda_tables=0.0,
                       prune_table_min_arity=3,
                       prune_min_stability=0.5,
                       prune_min_mean_gain=0.0,
                       prune_selector="ranked_path",
                       prune_path_steps=32,
                       prune_path_fraction=0.995,
                       prune_path_tolerance=0.001,
                       prune_main_effects=False,
                       band_tolerance=0.75,
                       band_deviance_cap=0.001,
                       prune_fold_min_rows=125,
                       prune_fold_es_patience=None,
                       prune_guard_min_rows=500,
                       prune_slope_eps=0.01,
                       prune_slope_min_z=3.0,
                       prune_size_penalty=None,
                       early_stopping=None,
                       unknown_category="rare")

Purpose¶

Implementation of the scikit-learn estimator API for t-boost classification. scikit-learn itself is optional: without it, the class runs standalone with the same methods.

Supports model training, inference and auxiliary calculations like rating tables, prediction contributions and feature importance.

The model depends on the number of classes in the target:

Two classes — A logistic model (see Classification: objectives and metrics).
Three or more classes — A softmax model with one set of rating tables per class (see Multiclassification: objectives and metrics).

Parameters¶

See Training parameters for the full list of parameters. n_trees, learning_rate and lambda_ can be passed by position; every other parameter is keyword-only. For multiclassification, see also Multiclassification settings.