Also known as: pool-adjacent-violators, PAV calibration, monotone calibration
Isotonic regression fits a non-parametric monotone function from raw scores to calibrated probabilities. More flexible than Platt scaling — handles any monotone miscalibration shape — at the cost of needing more labels and being prone to overfitting at the score-distribution tails.
Isotonic regression is the non-parametric counterpart to Platt scaling for score calibration . Instead of assuming a sigmoid shape, isotonic regression fits the most accurate monotone function from raw scores to empirical probabilities — a piecewise-constant step function with as many breakpoints as the data supports.
The fit is via the pool-adjacent-violators (PAV) algorithm. The result is the unique non-decreasing function
f* = argmin Σᵢ (f(sᵢ) − yᵢ)²
s.t. sᵢ ≤ sⱼ ⇒ f(sᵢ) ≤ f(sⱼ)Sort by raw score, walk forward, and whenever an interval’s mean violates monotonicity, merge it with the previous interval. The merged means form the calibration curve.
Platt is constrained to a two-parameter sigmoid shape — fast, data-efficient, but blind to miscalibration patterns the sigmoid family can’t represent. Isotonic regression has no such constraint: any monotone function is a candidate, and the algorithm picks the one that best matches the labels.
Isotonic regression buys you flexibility at the cost of data hunger. A two-parameter sigmoid fits on a few hundred labels; a non-parametric monotone function wants thousands to be stable, especially at the tails of the score distribution.
The principal failure mode of isotonic regression is overfitting at the top and bottom of the score distribution. Most of your labels live in the middle of the score range; the top few percent (the highest-scored documents) and the bottom few percent (the lowest-scored documents) are sparse. PAV merges those sparse bins into whatever step its few data points dictate — and that step is high-variance across resamples.
Mitigations:
A rough decision rule:
The other axis is how often you re-fit. Platt’s two parameters are nearly invariant to small label-set perturbations; isotonic’s step function moves around significantly each retraining. If you re-fit on a rolling window and want stable thresholds across windows, Platt has a steadiness advantage.
“Pool-adjacent-violators” describes the algorithm literally: walking through sorted points, whenever the current point’s running mean is less than the previous block’s mean (a “violator” of monotonicity), pool the two blocks together by averaging and continue. The name dates to Brunk (1955) and shows up unchanged in modern scikit-learn docs.
The algorithm is
Three things to know before shipping isotonic in a calibration pipeline:
Sort the data by raw score. Initialize each point as its own group with its own mean. Walk left-to-right; whenever an adjacent group has a smaller mean than the previous group (violates monotonicity), merge them and recompute the mean. Repeat until no violations remain. The merged-group means become the fitted calibration curve — a piecewise-constant monotone step function.
When the empirical calibration curve has structure a sigmoid cannot bend around — kinks, plateaus, or a non-symmetric S-shape. Modern neural-network reranker scores often show this because the model's score distribution is heavy-tailed and the relevance rate at the top of the distribution doesn't follow a smooth sigmoid.
Isotonic regression has very few labels at the high end of the score distribution (top-scored docs are rare), so the fitted curve there is determined by maybe a dozen points. The fit becomes a step function that snaps to whatever the empirical rate happens to be in those few bins — wildly noisy across draws. Mitigations include adding label smoothing, capping the fit's extreme bins, or falling back to Platt at the tails and isotonic in the middle.
