XGBoost Default Parameters

At the time of writing, the default XGBoost parameters are defined in the C code.

When the Python API specifies that a parameter has a None value, the default value from the C code is used.

Tree Booster Defaults

Default parameter values for the tree booster are defined in the file xgboost/src/tree/param.h.

The relevant default parameters are:

• learning_rate: 0.3
• min_split_loss: 0
• max_depth: 6
• max_leaves: 0
• max_bin: 256
• grow_policy: “depthwise”
• max_cat_to_onehot: 4
• max_cat_threshold: 64
• min_child_weight: 1
• reg_lambda: 1
• reg_alpha: 0
• max_delta_step: 0
• subsample: 1
• sampling_method: “uniform”
• colsample_bynode: 1
• colsample_bylevel: 1
• colsample_bytree: 1
• sketch_ratio: 2
• cache_opt: True
• refresh_leaf: True

Parameter aliases defined in this file include:

• reg_lambda, lambda
• reg_alpha, alpha
• min_split_loss, gamma
• learning_rate, eta

Linear Booster Defaults

Default parameter values for the linear booster are defined in the file xgboost/src/linear/param.h.

The relevant default parameters are:

• learning_rate: 0.5
• reg_lambda: 0
• reg_alpha: 0
• feature_selector: “cyclic”

Parameter aliases defined in this file include:

• reg_lambda, lambda
• reg_alpha, alpha
• learning_rate, eta

API Documentation

Default parameter values are also specified in the official API documentation.

The parameters and their defaults are as follows:

• booster: gbtree
• device: cpu
• verbosity: 1
• validate_parameters: false
• nthread: maximum number of threads available if not set
• disable_default_eval_metric: false
• eta: 0.3
• gamma: 0
• max_depth: 6
• min_child_weight: 1
• max_delta_step: 0
• subsample: 1
• sampling_method: uniform
• colsample_bytree: 1
• colsample_bylevel: 1
• colsample_bynode: 1
• lambda: 1
• alpha: 0
• tree_method: auto
• scale_pos_weight: 1
• refresh_leaf: 1
• process_type: default
• grow_policy: depthwise
• max_leaves: 0
• max_bin: 256
• num_parallel_tree: 1
• multi_strategy: one_output_per_tree
• max_cached_hist_node: 65536
• sample_type: uniform
• normalize_type: tree
• rate_drop: 0.0
• one_drop: 0
• skip_drop: 0.0
• objective: reg:squarederror
• seed: 0
• seed_per_iteration: false
• tweedie_variance_power: 1.5
• huber_slope: 1.0
• lambdarank_pair_method: mean
• lambdarank_unbiased: false
• lambdarank_bias_norm: 2.0
• ndcg_exp_gain: true
• save_period: 0
• task: train
• model_dir: models/
• dump_format: text
• name_dump: dump.txt
• name_pred: pred.txt
• pred_margin: 0