recommended_recipe¶
Return an estimator configured with the benchmark recipe.
The constructors of TBoostRegressor and TBoostClassifier already default to this recipe
(early stopping with adaptive patience, leaf refinement, bagging and pruning), so a bare
estimator is the benchmarked configuration. This function is the explicit entry point: it sets
the tree limit and the number of threads, and for the log-link objectives it also sets the
categorical encoding explicitly (cat_target="log_mean" for gamma) and reanchor=True.
Method call format¶
recommended_recipe(objective="squared_error",
*,
budget=4000,
n_jobs=4,
tuned=True,
seed=0,
**overrides)
Parameters¶
objective¶
Description¶
The objective. squared_error, poisson, gamma and tweedie return a TBoostRegressor;
logistic returns a TBoostClassifier (softmax is used automatically for three or more
classes).
Possible types
string
Default value
squared_error
budget¶
Description¶
The maximum number of trees (n_trees). Early
stopping decides the actual number.
Possible types
int
Default value
4000
n_jobs¶
Description¶
The number of threads (see n_jobs).
Possible types
int
Default value
4
tuned¶
Description¶
Apply the tuned part of the recipe: n_bags=8, colsample_bytree=0.8 and, for the log-link
objectives, reanchor=True, cat_target ("log_mean" for gamma, "mean" otherwise),
cat_leakage="kfold" and cat_k=5.
Possible types
bool
Default value
True
seed¶
Description¶
The random seed used for training.
Possible types
int
Default value
0
**overrides¶
Description¶
Any training parameter. These take precedence over the recipe.
Possible types
key=value format
Default value
None
Return value¶
An untrained TBoostRegressor or TBoostClassifier.