[default_model]

The model fitted for every outcome, and the optional recalibration applied to it afterwards. Each outcome gets its own fitted copy, so outcomes share these settings but not the fitted model.

[default_model]
algorithm = "HistGradientBoostingClassifier"

[default_model.hyperparameters]
learning_rate = 0.1
loss = "log_loss"
max_features = 0.2

[default_model.recalibration]
recalibrate = true
method = "isotonic"

[default_model.recalibration.hyperparameters]
y_min = 0.0
y_max = 1.0
out_of_bounds = "clip"

Key	Type	Default	Description
algorithm	string	required	Name of the model class. See available models.
hyperparameters	table	{}	Arguments passed to the model constructor.
recalibration	table	none	Recalibration applied after the model is fitted. Omit it to fit the model without recalibration.

Algorithm

algorithm is the exact class name of an estimator from scikit-learn (sklearn.ensemble, sklearn.linear_model, sklearn.isotonic), NGBoost or OrdBoost, or of a model you registered with ModelRegistry. The models page lists where names are looked up and how to register your own.

The name is looked up when the configuration is validated, so a misspelt name is reported before any data is loaded. A model registered with ModelRegistry is accepted if it is registered before the pipeline is created.

Hyperparameters

Every key under [default_model.hyperparameters] is passed to the constructor of the model. A key the model does not accept is reported when the model is created.

A hyperparameter can be a list of values to search over:

[default_model.hyperparameters]
learning_rate = [0.1, 0.05]
max_depth = 10

When grid_search = true in [workflow.validation.cross_validation] and the run mode runs cross-validation, every combination of the listed values is tried and the best model is kept. Scalar values are held fixed. Without a search, the first value of a list is used. See the run modes.

Recalibration

Recalibration adjusts the predicted probabilities of a fitted classifier so that they match the observed frequencies better. The model is first fitted on the training set, then a calibrator is fitted on a separate recalibration set.

Key	Type	Default	Description
recalibrate	boolean	required	Whether to recalibrate the fitted model.
method	string	"isotonic"	"isotonic" (isotonic regression), "sigmoid" (Platt scaling) or "temperature" (temperature scaling).
hyperparameters	table	{}	Reserved for hyperparameters of the recalibration method. Currently not used. See the note below.

Note

The hyperparameters of the recalibration are accepted by the configuration but are not used yet: they are not passed to the recalibration method, and any values are ignored. A future version may pass them directly to the recalibrator, for example y_min, y_max and out_of_bounds for the isotonic method. Until then, only recalibrate and method have an effect.

The recalibration set is defined by [workflow.validation.recalibration_split]. A recalibration table in the default model requires that split, even when recalibrate = false. To disable recalibration, either omit the recalibration table or set recalibrate = false and keep the split.