Objectives and metrics

This section contains basic information regarding the supported objectives for various machine learning problems.

• Regression
• Classification
• Multiclassification

Refer to the Variables used in formulas section for the description of commonly used variables in the listed metrics.

The objective is set by the objective parameter. Every objective is the deviance of a distribution with a link function, and the same deviance is used throughout the fit: the trees minimize it, early stopping monitors it on the validation dataset, and pruning compares sets of tables with it.

Objective	Machine learning problem	Link	Class
squared_error	regression	identity	TBoostRegressor
poisson	claim frequency, counts	log	TBoostRegressor
gamma	severity	log	TBoostRegressor
tweedie	pure premium	log	TBoostRegressor
logistic	binary classification	logit	TBoostClassifier
softmax (automatic for 3+ classes)	multiclassification	softmax	TBoostClassifier

Metrics can also be calculated separately from the training with the t_boost.metrics module: the deviances of the regression objectives and weight-aware ranking metrics (Gini, lift).