[outcome_overrides]

The default model is used for every outcome unless an outcome is overridden. An override is written as a table named after the outcome, for a column listed in data.outcomes:

[outcome_overrides.ANY_COMP]
algorithm = "GradientBoostingClassifier"

[outcome_overrides.ANY_COMP.hyperparameters]
learning_rate = [0.1, 0.05]
max_depth = 10

[outcome_overrides.ANY_COMP.recalibration]
recalibrate = false

An override has the same keys as [default_model], and all of them are optional. Only the keys that are written change the default model, so an override can, for example, change a single hyperparameter or only switch the recalibration off:

[outcome_overrides.ANY_COMP.hyperparameters]
max_depth = 5

[outcome_overrides.MORTALITY_30D.recalibration]
method = "sigmoid"

The name of an outcome that is not in data.outcomes is rejected. Outcomes without an override use the default model unchanged.

How an override is combined with the default

The settings of an override are applied on top of the default model, one group of settings at a time.

Algorithm and hyperparameters

• If the override uses the same algorithm as the default, or has no algorithm, its hyperparameters are merged into the default ones. Keys that are present in both take the override value, and the other default keys are kept.
• If the override uses a different algorithm, the default hyperparameters are discarded and only the hyperparameters written in the override are used. This avoids passing arguments that the new model does not accept.

For example, with the default model above:

Override	Resulting hyperparameters
algorithm = "HistGradientBoostingClassifier", learning_rate = 0.5	learning_rate = 0.5, loss = "log_loss", max_features = 0.2
algorithm = "RandomForestClassifier", n_estimators = 50	n_estimators = 50
algorithm = "RandomForestClassifier"	none, so the model defaults are used

Recalibration

• If the override has no recalibration table, the default recalibration is kept, even when the algorithm changes.
• If the override has the same method as the default, or no method, the keys it sets replace the default ones. Its hyperparameters are merged into the default ones.
• If the override has a different method, the default recalibration hyperparameters are discarded and only those written in the override are used. recalibrate is kept from the default unless the override sets it.
• If the default model has no recalibration, an override that adds one must set recalibrate.
• Setting recalibrate = false in an override turns the recalibration off for that outcome.

In the example above, ANY_COMP uses a different algorithm and disables recalibration. MORTALITY_30D has no override and uses the default model with isotonic recalibration.

Whenever the default model has a recalibration table, or an override turns recalibration on for an outcome, [workflow.validation.recalibration_split] must be defined.