Module Modelkit.Multiclass_ranking

Ranking metrics over class-probability matrices.

One-versus-rest ROC AUC scores every class column against its indicator; Macro averages classes equally, Weighted by weighted truth support, and Micro pools every row-class indicator into one binary curve. One-versus-one ROC AUC averages, over every pair of classes with positive weighted support in ascending order, the mean of the two directional AUCs on the rows belonging to that pair; Weighted uses the pair's weighted prevalence, and Micro is a typed validation error. A class without weighted support follows the undefined policy with a fallback of 0.5. scikit-learn refuses sample weights for one-versus-one AUC; ModelKit applies them to both the pairwise curves and the prevalences.

Top-k accuracy counts a row as a hit when fewer than k other classes outrank the truth class, with exact ties broken toward the higher column index as scikit-learn does. k must lie in [1, classes). One-versus-rest costs O(classes * n log n); one-versus-one costs O(classes squared * n log n); top-k costs O(n * classes).