Metrics 0.3.0 multilabelconfusionmatrix

Lodestar.Metrics 0.3.0. 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.

MultilabelConfusionMatrix

One 2×2 ConfusionMatrix per label — each label counted against everything else, which is what a per-class score is built from before it is divided into one.

A stack of the existing type, not a new one. Each entry is a two-label matrix, so everything that reads one reads these: Recall.Score over the entry for class 1 gives class 1's recall, and a test asserts exactly that. Wrapping the array in a type of its own would have bought a name and cost that.

Each entry's labels are 0 and 1 in that order, which puts the cells where the reference puts them: [0, 0] true negative, [0, 1] false positive, [1, 0] false negative, [1, 1] true positive.

Per label, or per sample

samplewise turns the count on its side: one matrix per row, counting that row's labels rather than each label's samples. On two samples over three labels it returns two matrices instead of three.

The reference offers it on a label matrix only, and refuses it on single-label input with "Samplewise metrics are not available outside of multilabel classification". Here that refusal is structural rather than checked — samplewise is a parameter of the matrix overload and does not exist on the other, so the call a caller must not make cannot be written.

Members

Member	What it does
MultilabelConfusionMatrix.Compute	The stack, from labels or from a matrix.