Fits a logistic regression model to recalibrate probabilities.
Parameters should be estimated on a held-out calibration set.
Usage
fit_platt_scaling(probs, labels)
Arguments
- probs
Calibration set predicted probabilities
- labels
Calibration set labels
Value
A function that recalibrates probabilities
Details
Platt scaling fits: P(y=1|f) = 1 / (1 + exp(A*f + B))
where f is the original score/probability and A, B are fitted parameters.
Examples
if (FALSE) { # \dontrun{
# Fit on calibration set
cal_probs <- runif(100)
cal_labels <- rbinom(100, 1, cal_probs)
calibrator <- fit_platt_scaling(cal_probs, cal_labels)
# Apply to new predictions
new_probs <- c(0.2, 0.5, 0.8)
calibrated <- calibrator(new_probs)
} # }