Metrics and Optimization Cost

Users choose a human-facing metric. Internally, every strategy minimizes a cost:

Metric	Task	Better	Internal cost
RMSE, MAE, log loss	Applicable task	Lower	Metric value
R2	Regression	Higher	-R2
Accuracy, ROC AUC, PR AUC, macro F1	Classification	Higher	1 - metric

This explains why the best metric can rise while the optimization cost falls. The dashboard emphasizes the selected metric and formats cost as an optimizer diagnostic rather than a competing result.

Canonical metric IDs are rmse, mae, r2, accuracy, roc_auc, pr_auc, log_loss, and f1_macro. Short aliases are not accepted in 1.0.

CV optimization uses the mean fold cost and preserves each fold's metric plus the population standard deviation. positive_label explicitly chooses the binary positive class; sensitivity, specificity and thresholded predictions use that class. AUC and PR AUC use its scores. Residuals are actual minus predicted.

Objective validation checks task and target domains for every objective in a search space. For example, Gamma targets must be positive, Poisson and Tweedie targets must be nonnegative with a positive value, and squared-log targets must exceed -1. The supported objective definitions follow the XGBoost parameters and LightGBM parameters.