2.3. Boosting Algorithms Custom Cost Functions

Empulse provides a number of cost functions for use with the xgboost.XGBClassifier. These cost function wrap around various implementation of the expected_cost_loss. They compute the gradient and hessian of the expected cost loss, which is then used by the boosting algorithm to optimize the cost function.

The empulse.models module defines the empulse.models.B2BoostClassifier and empulse.models.CSBoostClassifier as a convenience, but you easily define your own XGBoost model and use the cost functions directly.

For example, to use the empulse.metrics.make_objective_aec with xgboost, you can do the following:

import xgboost as xgb
from sklearn.datasets import make_classification
from empulse.metrics import make_objective_aec

X, y = make_classification()
cost_function = make_objective_aec(fp_cost=5, fn_cost=1, tp_cost=1, tn_cost=1)
clf = xgb.XGBClassifier(objective=cost_function, n_estimators=100, max_depth=3)
clf.fit(X, y)

Empulse also contains empulse.metrics.make_objective_churn and empulse.metrics.make_objective_acquisition functions for optimizing case-specific implementation of the expected_cost_loss.