Fits isotonic (monotonic) regression for probability calibration.
Non-parametric method that produces a monotonically increasing
calibration mapping.
Usage
fit_isotonic_calibration(probs, labels)
Arguments
- probs
Calibration set predicted probabilities
- labels
Calibration set labels
Value
A function that recalibrates probabilities
Details
Isotonic regression finds the best monotonically increasing function
that maps scores to calibrated probabilities. It's more flexible than
Platt scaling but requires more calibration data.
IMPORTANT: Must be fit on a held-out calibration set, not training data,
to avoid leakage.
Examples
if (FALSE) { # \dontrun{
# Fit on calibration set
calibrator <- fit_isotonic_calibration(cal_probs, cal_labels)
# Apply to new predictions
calibrated <- calibrator(new_probs)
} # }