check_bindings

Raise an error if a parameter set away from its default had no effect on the fit.

This is the strict counterpart of the binding_report_ attribute, for a model whose parameters are part of a filed or otherwise defended model, where a silently ignored setting is a defect.

Method call format

check_bindings()

Return value

None. Raises ValueError listing every parameter whose status in binding_report_ is INERT (it had no effect on this fit) or OVERRIDDEN (another parameter's value voided it).

binding_report_

One entry per checked parameter set away from its default, with param, value, status (HONOURED, INERT or OVERRIDDEN), reason and overridden_by.

The report covers parameters whose effect depends on other settings: the pruning gates and limits (for example, every pruning parameter when prune=False, and the parameters of the pruning selector that did not run), validation_fraction with an eval_set, monotone_constraints and deprecated spellings. It is not an audit of every parameter.

Usage examples

from t_boost import TBoostRegressor

model = TBoostRegressor(objective="poisson", prune_table_budget=4)
model.fit(train_data, "ClaimCount", exposure="Exposure")
for row in model.binding_report_:
    print(row["param"], row["status"], row["reason"])
model.check_bindings()