Skip to content

API reference — Evaluation reports

What Metrics returns, field by field — the contract every learner is scored under. Reports are plain dicts; missing pairs are dropped from the math and counted in n_missing, never hidden.

Metrics().regression(y_true, y_pred) → dict

field meaning
n, n_missing observed pairs; dropped missing pairs
mean_error mean(actual − fitted) — signed
bias mean_pred / mean_true (1.0 = unbiased)
mae, mad mean / median absolute deviation
mape absolute % error, non-zero targets only
rmse, r2, max_error the usual
explained_variance 1 − Var(error)/Var(y)
deciles per-target-decile error profile — the DAR "bathtub" view

Metrics().classification(y_true, y_pred, probabilities=None) → dict

field meaning
n, n_missing, n_classes observed pairs; classes seen
accuracy overall accuracy
macro_precision/recall/f1 unweighted mean over classes
per_class precision/recall/f1/support per class
confusion counts matrix
auc, logloss, brier only with probabilities (P(class 1))

Grouped reports

classification_by_group(keys, ...) / regression_by_group(keys, y_true, y_pred) return one report per group, as a dict keyed by the group label — the format ModelTree's group API consumes.

Real values for every field on real data: the Metrics & plots showcase and the Metrics page example.

Full generated listing · next: Visualization