User request

While working on scikit-learn#26120, I found that roc_curve can return thresholds greater than 1. A non-regression test could be as follows:

import numpy as np
from sklearn.metrics import roc_curve

def test_roc_curve_with_probablity_estimates():
    rng = np.random.RandomState(42)
    y_true = rng.randint(0, 2, size=10)
    y_score = rng.rand(10)
    _, _, thresholds = roc_curve(y_true, y_score)
    assert np.logical_or(thresholds <= 1, thresholds >= 0).all()

This is due to prepending a sentinel threshold in sklearn/metrics/_ranking.py to add the (fpr=0, tpr=0) point; the current logic uses thresholds[0] + 1, which can push the first threshold above 1 even when y_score are probabilities in [0, 1].

Reference (upstream): https://github.com/scikit-learn/scikit-learn/blob/e886ce4e1444c61b865e7839c9cff5464ee20ace/sklearn/metrics/_ranking.py#L1086

Researcher specification (proposed fix)

• Root cause: roc_curve prepends a threshold strictly greater than the maximum score to ensure the ROC starts at (0,0). In the current code this is thresholds = np.r_[thresholds[0] + 1, thresholds]. For probability-like scores in [0,1], this yields a threshold > 1.

• Proposed change (preserve semantics, minimize out-of-range values):

  • For float scores (probabilities), use the next representable float above max(y_score) via np.nextafter (dtype-aware). This keeps the threshold strictly greater than any score while minimally exceeding the domain when max == 1.0.
  • For integer scores, keep the current behavior: max + 1.

  Pseudocode:

  # File: sklearn/metrics/_ranking.py, in roc_curve
  if np.issubdtype(thresholds.dtype, np.floating):
      prepend_thresh = np.nextafter(thresholds[0], np.inf, dtype=thresholds.dtype)
  else:
      prepend_thresh = thresholds[0] + 1
  thresholds = np.r_[prepend_thresh, thresholds]

• Rationale: Using np.nextafter preserves the mapping invariant predicted = y_score >= thresholds[i] and keeps thresholds strictly decreasing, while avoiding large jumps above 1 for probability scores. Clipping to 1.0 would break semantics when max(y_score) == 1.0.

• Tests to add (in sklearn/metrics/tests/test_ranking.py):

  1. Probabilities with max < 1: thresholds remain within [0,1], first threshold is > max(y_score).
  2. Probabilities with an exact 1.0: first threshold is np.nextafter(1.0, +inf); mapping semantics intact.
  3. Non-probability (e.g., decision function): thresholds may exceed 1; first threshold is strictly greater than max(score); monotonic decrease and shapes preserved.

• Docstring tweak: clarify that thresholds[0] is a sentinel strictly greater than max(y_score); for float scores we use the next representable float; for integers, max + 1.

Acceptance

• Implement the above change and add the non-regression tests.
• Validate locally by running only the affected tests.
• Open a PR targeting branch scikit-learn__scikit-learn-26194.