src/eval/metrics.py

"""Detection + fairness metrics following the paper's definitions."""
from __future__ import annotations

from typing import Dict, List, Sequence

import numpy as np
from sklearn.metrics import roc_auc_score, accuracy_score, roc_curve


def _tpr_at_fpr(y_true: np.ndarray, y_score: np.ndarray, fpr_target: float) -> float:
    fpr, tpr, _ = roc_curve(y_true, y_score)
    if (fpr <= fpr_target).any():
        return float(tpr[fpr <= fpr_target].max())
    return 0.0


def compute_detection_metrics(y_true, y_score) -> Dict[str, float]:
    y_true = np.asarray(y_true); y_score = np.asarray(y_score)
    try:
        auc = roc_auc_score(y_true, y_score)
    except ValueError:
        auc = float("nan")
    acc = accuracy_score(y_true, (y_score > 0.5).astype(int))
    t1 = _tpr_at_fpr(y_true, y_score, 0.01)
    t01 = _tpr_at_fpr(y_true, y_score, 0.001)
    return {"auc": auc, "acc": acc, "tpr@fpr=1%": t1, "tpr@fpr=0.1%": t01}


def compute_fairness_metrics(
    y_true: Sequence[int],
    y_score: Sequence[float],
    groups: Sequence[str],
) -> Dict[str, float]:
    """F_FPR, F_MEO, F_DP, F_OAE — stdev-type measures across groups."""
    y_true = np.asarray(y_true); y_score = np.asarray(y_score)
    groups = np.asarray(groups)

    def _group_fpr(g):
        m = (groups == g) & (y_true == 0)
        if m.sum() == 0: return 0.0
        return float(((y_score > 0.5)[m]).mean())

    def _group_tpr(g):
        m = (groups == g) & (y_true == 1)
        if m.sum() == 0: return 0.0
        return float(((y_score > 0.5)[m]).mean())

    def _group_acc(g):
        m = (groups == g)
        if m.sum() == 0: return 0.0
        return float(((y_score > 0.5)[m] == y_true[m]).mean())

    def _group_dp(g):
        m = (groups == g)
        if m.sum() == 0: return 0.0
        return float((y_score > 0.5)[m].mean())

    uniq = sorted(set(groups.tolist()))
    fprs = [_group_fpr(g) for g in uniq]
    tprs = [_group_tpr(g) for g in uniq]
    accs = [_group_acc(g) for g in uniq]
    dps = [_group_dp(g) for g in uniq]

    def _std(xs): return float(np.std(xs)) * 100  # paper reports in %

    return {
        "F_FPR": _std(fprs),
        "F_MEO": max(max(fprs) - min(fprs), max(tprs) - min(tprs)) * 100,
        "F_DP": _std(dps),
        "F_OAE": _std(accs),
        "group_fprs": {g: v for g, v in zip(uniq, fprs)},
        "group_accs": {g: v for g, v in zip(uniq, accs)},
    }