balanced_accuracy.py

import numpy as np
import evaluate
import datasets

_DESCRIPTION = """
Balanced Accuracy for imbalanced classification.

Definitions
- Binary: (TPR + TNR) / 2
- Multiclass: macro-average of per-class recall

Extras
- threshold="auto": pick the best threshold for binary probabilities (Youden's J)
- ignore_index: skip unlabeled samples (e.g., -100)
- adjusted=True: sklearn-style chance correction
- return_per_class=True: also return per-class recalls (multiclass)
- class_mask=[...] (multiclass): average over a subset of classes
- support_per_class: when return_per_class=True (multiclass), also return true sample counts per class
- sample_weight: per-sample weights for binary/multiclass; replaces counts with weighted sums
"""

_KWARGS_DESCRIPTION = """
Args:
    predictions: 1D list/array.
        Binary: integer labels {0,1}, or probabilities in [0,1] (if threshold given)
        Multiclass: integer labels {0..K-1}
    references:  1D list/array of integer labels.
    task:        "binary" | "multiclass" (Default: "binary")
    num_classes: int, for multiclass; inferred if labels are 0..K-1.
    adjusted:    bool, default False. (binary: 2*BA-1; multiclass: (BA-1/K)/(1-1/K))
    zero_division: float, default 0.0.
    threshold:   float in (0,1) or "auto" (binary only).
    ignore_index: int | None. If set, samples with reference == ignore_index are skipped.
    return_per_class: bool, default False — also return per-class recalls list (multiclass).
    class_mask: Optional[list[int]] — only average over these classes (multiclass).
    sample_weight: Optional[list[float]] — per-sample weights.
Returns:
    {"balanced_accuracy": float}
    + (binary, threshold="auto"): {"optimal_threshold": float}
    + (multiclass, return_per_class=True):
        {"per_class_recall": list[float], "support_per_class": list[int or float]}
"""

_CITATION = ""


def _safe_div(num, den, zero_div=0.0):
    num = np.asarray(num, dtype=float)
    den = np.asarray(den, dtype=float)
    out = np.full_like(num, float(zero_div))
    mask = den != 0
    out[mask] = num[mask] / den[mask]
    return out

def _is_integer_like(arr, atol=1e-12):
    """Return True if all values are finite and very close to integers."""
    arr = np.asarray(arr, dtype=float)
    if not np.all(np.isfinite(arr)):
        return False
    return np.all(np.abs(arr - np.round(arr)) <= atol)


def _check_1d_same_len(y_true, y_pred, name_true="references", name_pred="predictions"):
    if y_true.ndim != 1 or y_pred.ndim != 1:
        raise ValueError(f"`{name_true}` and `{name_pred}` must be 1D.")
    if y_true.shape[0] != y_pred.shape[0]:
        raise ValueError(f"Length mismatch: `{name_true}`={y_true.shape[0]} vs `{name_pred}`={y_pred.shape[0]}.")
    if not np.all(np.isfinite(y_pred)):
        raise ValueError("`predictions` contains NaN/Inf.")


def _binary_ba_from_labels(y_true_i, y_pred_i, zero_div):
    tp = float(((y_true_i == 1) & (y_pred_i == 1)).sum())
    fn = float(((y_true_i == 1) & (y_pred_i == 0)).sum())
    tn = float(((y_true_i == 0) & (y_pred_i == 0)).sum())
    fp = float(((y_true_i == 0) & (y_pred_i == 1)).sum())
    tpr = _safe_div(tp, tp + fn, zero_div)
    tnr = _safe_div(tn, tn + fp, zero_div)
    return 0.5 * (tpr + tnr)


def _binary_find_best_threshold(y_true, probs, zero_div):
    p = np.asarray(probs, dtype=float)
    uniq = np.unique(p)
    if uniq.size == 1:
        candidates = [float(uniq[0])]
    else:
        mids = (uniq[:-1] + uniq[1:]) / 2.0
        candidates = [float(uniq[0] - 1e-12), *mids.tolist(), float(uniq[-1] + 1e-12)]

    best_t, best_ba = None, -1.0
    yt = (np.asarray(y_true) == 1).astype(int)
    for t in candidates:
        yp = (p >= t).astype(int)
        ba = _binary_ba_from_labels(yt, yp, zero_div)
        if (ba > best_ba) or (abs(ba - best_ba) < 1e-12 and (best_t is None or abs(t - 0.5) < abs(best_t - 0.5))):
            best_ba, best_t = float(ba), float(t)
    return best_t, best_ba


class BalancedAccuracy(evaluate.Metric):
    def _info(self):
        return evaluate.MetricInfo(
            description=_DESCRIPTION,
            citation=_CITATION,
            inputs_description=_KWARGS_DESCRIPTION,
            features=datasets.Features(
                {"predictions": datasets.Value("float64"),
                 "references":  datasets.Value("float64")}
            ),
            reference_urls=[
                "https://scikit-learn.org/stable/modules/generated/sklearn.metrics.balanced_accuracy_score.html"
            ],
        )

    def _compute(
        self,
        predictions,
        references,
        task: str = "binary",
        num_classes: int | None = None,
        adjusted: bool = False,
        zero_division: float = 0.0,
        threshold: float | str | None = None,
        ignore_index: int | None = None,
        return_per_class: bool = False,
        class_mask: list[int] | None = None,
        sample_weight: list[float] | None = None,
    ):
        y_true_all = np.asarray(references).astype(int)
        y_pred_all = np.asarray(predictions)

        # ignore_index mask
        if ignore_index is not None:
            mask = y_true_all != ignore_index
        else:
            mask = np.ones_like(y_true_all, dtype=bool)

        y_true = y_true_all[mask]
        y_pred_in = y_pred_all[mask]

        if y_true.size == 0:
            return {"balanced_accuracy": float("nan"), "reason": "empty_after_ignore_index"}

        # weights
        w = None
        if sample_weight is not None:
            w_in = np.asarray(sample_weight, dtype=float)
            if w_in.shape[0] != y_true_all.shape[0]:
                raise ValueError("`sample_weight` length must match number of samples.")
            w = w_in[mask]

        _check_1d_same_len(y_true, y_pred_in)

        # ---- Binary ----
        if task == "binary":
            uniq_pred = np.unique(y_pred_in)

            if np.isin(uniq_pred, [0.0, 1.0]).all():
                y_pred = y_pred_in.astype(int)

            elif _is_integer_like(y_pred_in):
                raise ValueError("For binary with label predictions, values must be 0/1.")
            else:
                if np.any((y_pred_in < 0) | (y_pred_in > 1)):
                    raise ValueError("For binary with probabilities, `predictions` must be in [0,1].")
                if threshold == "auto":
                    t_opt, ba = _binary_find_best_threshold(y_true, y_pred_in, zero_division)
                    if adjusted:
                        ba = 2 * ba - 1
                    return {"balanced_accuracy": float(ba), "optimal_threshold": float(t_opt)}
                else:
                    t = 0.5 if (threshold is None) else float(threshold)
                    if not (0.0 < t < 1.0):
                        raise ValueError("`threshold` must be in (0,1) or 'auto'.")
                    y_pred = (y_pred_in >= t).astype(int)

            # weighted confusion
            if w is None:
                tp = float(((y_true == 1) & (y_pred == 1)).sum())
                fn = float(((y_true == 1) & (y_pred == 0)).sum())
                tn = float(((y_true == 0) & (y_pred == 0)).sum())
                fp = float(((y_true == 0) & (y_pred == 1)).sum())
            else:
                tp = float(w[((y_true == 1) & (y_pred == 1))].sum())
                fn = float(w[((y_true == 1) & (y_pred == 0))].sum())
                tn = float(w[((y_true == 0) & (y_pred == 0))].sum())
                fp = float(w[((y_true == 0) & (y_pred == 1))].sum())

            tpr = _safe_div(tp, tp + fn, zero_division)
            tnr = _safe_div(tn, tn + fp, zero_division)
            ba = 0.5 * (tpr + tnr)
            if adjusted:
                ba = 2 * ba - 1
            return {"balanced_accuracy": float(ba)}

        # ---- Multiclass ----
        if task != "multiclass":
            raise ValueError("`task` must be 'binary' or 'multiclass'.")

        y_pred = y_pred_in.astype(int)

        if num_classes is None:
            num_classes = int(max(y_true.max() if y_true.size else 0,
                                  y_pred.max() if y_pred.size else 0)) + 1
        if num_classes <= 0:
            raise ValueError("`num_classes` must be positive.")
        if (y_pred < 0).any() or (y_pred >= num_classes).any():
            raise ValueError(f"`predictions` must be in [0,{num_classes-1}] for multiclass.")
        if (y_true < 0).any() or (y_true >= num_classes).any():
            raise ValueError(f"`references` must be in [0,{num_classes-1}] for multiclass.")

        classes = list(range(num_classes))
        if class_mask is not None and len(class_mask) > 0:
            classes = [c for c in class_mask if 0 <= c < num_classes]
            if len(classes) == 0:
                return {"balanced_accuracy": float("nan"), "reason": "empty_class_mask_after_filtering"}

        recalls, supports = [], []
        for c in classes:
            mask_c = (y_true == c)
            if w is None:
                denom = float(mask_c.sum())
                num = float((mask_c & (y_pred == c)).sum())
                supports.append(int(denom))
            else:
                denom = float((w[mask_c]).sum())
                num = float((w[mask_c & (y_pred == c)]).sum())
                supports.append(float(denom))
            recalls.append(float(_safe_div(num, denom, zero_division)))

        recall_c = np.asarray(recalls, dtype=float)
        ba = float(recall_c.mean())
        if adjusted:
            chance = 1.0 / float(len(classes))
            ba = float((ba - chance) / (1.0 - chance))

        out = {"balanced_accuracy": ba}
        if return_per_class:
            out["per_class_recall"] = recall_c.tolist()
            out["support_per_class"] = supports
        return out