Skip to main content
Back

Researchers Shrink Bloated AI Confidence Regions With a Single Geometric Score

A new non-conformity score based on the distance to the median-ranked neighbor makes conformal prediction regions immune to nearly half-sample contamination.

Paper, Full Access
  • New conformal score uses the distance to the (n/2+1)-th nearest neighbor, robust to nearly half-sample contamination.
  • Score is equivalent to the smallest ball radius covering a strict majority of training points around a candidate.
  • Prediction region provably converges to a population level set of the distance-to-a-measure functional.
  • Exponential concentration and Hausdorff-distance bounds are derived at rate sqrt(log n / n).
  • Empirical region equals points covered by at least k balls of radius beta centered on the data.
  • Cheap inner approximation from per-point local k-NN radii avoids grid search entirely.

Conformal prediction has become the standard recipe for wrapping a distribution-free confidence region around any predictor. The guarantee only says the true point lands inside the region with the right frequency, though; it says nothing about the shape of the region itself. A new paper from Cholaquidis, Joly, and Moreno attacks that weakness by swapping the usual residual-based score for a geometric one that shrugs off outliers by construction.

The pitch is simple. Instead of scoring how atypical a point is by how far it deviates from a regression fit, score it by the radius of the smallest ball centered on it that captures at least half of the training sample. The non-conformity score becomes the half-mass radius around a point, equivalently the distance to its (n/2+1)-nearest neighbour. Up to nearly half the data can drift arbitrarily far away and the score barely moves.

Why residual scores produce bloated regions

Standard conformal prediction inherits a coverage guarantee under exchangeability regardless of the score function, but the score controls the geometry. When the data is multi-modal, residual-based scores tend to yield prediction sets that are unnecessarily large, easily swayed by a small fraction of atypical points, and poorly aligned with the underlying distribution. With heavy tails or clusters, a handful of extreme points can balloon the region until it wraps a lot of empty space.