Lodestar.Conformal 0.1.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

The prediction set: every class whose probability clears 1 − q. Sometimes that is none of them.

public static bool[] PredictionSet(ReadOnlySpan<double> probabilities, double quantile)

Parameters — probabilities is one sample's predicted probabilities, in the same class order LeastAmbiguousScores was given. quantile is the calibrated quantile from Quantile.

Returns — a fresh bool[] of the same length, true where that class is in the set.

Exceptions — ArgumentOutOfRangeException when quantile is negative or NaN.

Example — a calibrated quantile of 0.5, so the threshold is 0.5. One confident row keeps one class; one undecided row keeps none.

using Lodestar.Conformal;

bool[] confident = SplitConformal.PredictionSet([0.75, 0.15, 0.10], 0.5);
bool firstIn = confident[0];    // => True
bool secondIn = confident[1];   // => False

bool[] undecided = SplitConformal.PredictionSet([0.40, 0.35, 0.25], 0.5);
bool anyIn = Array.Exists(undecided, included => included);   // => False

Remarks — the empty set is a real answer, and it is not repaired here. When no class clears the threshold, LAC says so, and that is information: the model is less sure about this sample than it was about 1 − alpha of the calibration set. Substituting the most likely class would return something with no coverage guarantee under a name that promises one — the same mistake as clamping the quantile, which decision 0070 refuses for the same reason. If your call site must produce a class, take the arg-max yourself, knowingly.

A set with two or more classes is the other half of the same signal, and it is the usual reason to reach for conformal classification at all: the model is telling you which alternatives it could not rule out at this level. An infinite quantile returns every class, which is the trivial prediction the calibration size forced.

Coverage is a statement about the calibration set as a whole, not about this row: 1 − alpha of exchangeable samples have their true class in the set. Nothing says which ones.

The guarantee assumes exchangeability — see the guide's Exchangeability section.

Applies to — net10.0, netstandard2.0.

See also — SplitConformal.LeastAmbiguousScores, SplitConformal.Quantile, the Python equivalence table.