import numpy as np from sklearn.metrics import average_precision_score, precision_recall_curve, roc_auc_score from .decision import expected_cost def recall_at_precision(y_true, proba, min_precision: float) -> float: """Best recall achievable while keeping precision >= min_precision (0.0 if unattainable).""" precision, recall, _ = precision_recall_curve(y_true, proba) ok = precision >= min_precision return float(recall[ok].max()) if ok.any() else 0.0 def confusion(y_true, flagged) -> dict: y = np.asarray(y_true).astype(int) f = np.asarray(flagged).astype(bool) return { "tn": int(((y == 0) & ~f).sum()), "fp": int(((y == 0) & f).sum()), "fn": int(((y == 1) & ~f).sum()), "tp": int(((y == 1) & f).sum()), } def point_metrics(y_true, proba, threshold, cost_fn, cost_fp, min_precision=0.8) -> dict: y = np.asarray(y_true).astype(int) p = np.asarray(proba, dtype=float) flagged = p >= threshold cm = confusion(y, flagged) tp, fp, fn = cm["tp"], cm["fp"], cm["fn"] precision = tp / (tp + fp) if tp + fp else 0.0 recall = tp / (tp + fn) if tp + fn else 0.0 f1 = 2 * precision * recall / (precision + recall) if precision + recall else 0.0 has_both = 0 < y.sum() < len(y) return { "pr_auc": float(average_precision_score(y, p)) if has_both else float("nan"), "roc_auc": float(roc_auc_score(y, p)) if has_both else float("nan"), "recall_at_precision": recall_at_precision(y, p, min_precision) if has_both else float("nan"), "precision": precision, "recall": recall, "f1": f1, "accuracy": (cm["tn"] + tp) / len(y), "expected_cost": expected_cost(y, flagged, cost_fn, cost_fp), } def bootstrap_ci(y_true, proba, threshold, cost_fn, cost_fp, min_precision=0.8, n_boot=1000, ci=0.95, seed=0) -> dict: """Percentile bootstrap over transactions. Returns {metric: [lo, hi]}.""" y = np.asarray(y_true).astype(int) p = np.asarray(proba, dtype=float) rng = np.random.default_rng(seed) samples = {} n = len(y) for _ in range(n_boot): idx = rng.integers(0, n, n) if y[idx].sum() == 0 or y[idx].sum() == n: continue m = point_metrics(y[idx], p[idx], threshold, cost_fn, cost_fp, min_precision) for k, v in m.items(): samples.setdefault(k, []).append(v) lo_q, hi_q = (1 - ci) / 2 * 100, (1 + ci) / 2 * 100 return {k: [float(np.percentile(v, lo_q)), float(np.percentile(v, hi_q))] for k, v in samples.items()}