fraud_fixed_fpr.py

"""Synthetic fraud-review example under a fixed false-positive-rate cap."""

from __future__ import annotations

import json
from dataclasses import asdict

import numpy as np

from class_imbalance.operating_points import operating_point_at_fpr_cap


def run_fraud_review_demo(random_state: int = 42) -> dict[str, float]:
    """Select an empirical fraud-review threshold under a one-percent FPR cap."""

    rng = np.random.default_rng(random_state)
    y_true = rng.binomial(1, 0.02, size=5_000)
    y_score = rng.beta(1.5, 12.0, size=y_true.size)
    y_score[y_true == 1] = rng.beta(4.0, 3.0, size=int(np.sum(y_true)))

    point = operating_point_at_fpr_cap(y_true, y_score, fpr_cap=0.01)
    reviewed = y_score >= point.threshold
    result = {key: float(value) for key, value in asdict(point).items()}
    result["review_rate"] = float(np.mean(reviewed))
    return result


def main() -> None:
    print(json.dumps(run_fraud_review_demo(), ensure_ascii=False, indent=2))


if __name__ == "__main__":
    main()