src/coredd/metrics.py

"""Metrics for the benchmark, with the sklearn calling convention.

Every metric is called ``metric(y_true, y_pred)`` on 1-D arrays over candidate nodes --
by default the nodes of every change pooled together, with ``pooled=False`` the nodes
of one change at a time -- so any scikit-learn classification metric drops in unchanged
and CoReDD stays metric-agnostic. This module only keeps what scikit-learn does not
provide: the ``rank`` marker, ``mrr``, and an ``auroc`` wrapper that is NaN where
``roc_auc_score`` raises, so a degenerate change is dropped from a per-case average
rather than aborting the run.

A metric is either **label** (``y_pred`` is boolean: which nodes were predicted) or
**rank** (``y_pred`` is a per-node score). The ``rank`` marker tells the evaluator to
hand the metric the node scores rather than a boolean mask. Pass ``sklearn.metrics``
functions directly for label metrics; they are treated as label by default.
"""

from __future__ import annotations

import numpy as np
from sklearn.metrics import roc_auc_score


def rank(func):
    """Mark a metric as consuming per-node scores rather than a boolean mask."""
    func.rank = True
    return func


def is_rank(metric) -> bool:
    """Whether a metric expects scores (``rank``) instead of a boolean prediction."""
    return bool(getattr(metric, "rank", False))


@rank
def mrr(y_true, scores) -> float:
    """Reciprocal rank of the first true positive under the score order; NaN if none."""
    y_true = np.asarray(y_true, dtype=bool)
    scores = np.asarray(scores, dtype=float)
    order = np.argsort(-scores, kind="stable")
    ranked = y_true[order]
    if not ranked.any():
        return float("nan")
    return 1.0 / (int(np.argmax(ranked)) + 1)


@rank
def auroc(y_true, scores) -> float:
    """ROC AUC; NaN when there is no positive or no negative to separate."""
    y_true = np.asarray(y_true, dtype=bool)
    if not y_true.any() or y_true.all():
        return float("nan")
    return float(roc_auc_score(y_true, scores))