simulate_decision

Scope: compute · Signature: simulate_decision(experiment_id: str, kind: {business_value|multiclass_cost|decision_layer|ranking_at|reco_business_value|ltr_business_value|unsupervised_threshold}, params: dict[str, Any] | None = None) -> Any

Use to choose an operating point: what threshold / cost / capacity yields.

kind="business_value" (binary): value, F-beta and capacity trade-offs; params cost_fp, cost_fn, gain_tp, gain_tn, fbeta, capacity_top_k | capacity_fraction, objective, compare_experiment_id (same workspace). Capacity selects a sampled threshold, not an exact quota; ties stay together. at_capacity reports its yield, possibly below budget; infeasible caps are refused. kind="multiclass_cost": params cost_matrix keyed by class label, {truth: {prediction: cost}}, every class in each row (diagonal too). Both take split (default validation: choose there, read test once). kind="decision_layer": calibration / threshold candidates (params methods). kind="ranking_at": at a popularity discount (alpha, value_weights, value_alpha). kind="reco_business_value" / "ltr_business_value" (+ target_key). kind="unsupervised_threshold": the anomaly threshold curve (n_points), plus current_point at the exact active cutoff and tie rule; label metrics are in-sample. Nothing is applied: recommend a point; a person applies it in the UI.

Powered by: mli decision support (mli/docs/guides/decision-support.md).