Metrics 0.3.0 topkaccuracy

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.

TopKAccuracy

Classification wearing a ranking's clothes. The input is a score per class per sample — the same shape a multiclass classifier's predict_proba produces — and a sample counts as correct when its true class is anywhere among the k highest-scoring, rather than only when it is the single highest. At k = 1 it is ordinary accuracy.

It is on this page rather than with the classification metrics because what it measures is a position in an ordering, and because it shares the tie rule the rest of the page uses: equal scores are ranked in descending index order, which is what scikit-learn's stable sort gives. A tie straddling the k boundary therefore has a determined answer, not an arbitrary one.

One divergence from top_k_accuracy_score, and it is a widening rather than a narrowing. scikit-learn infers the class set from y_true and refuses a score row wider than what it found unless it is given labels; here the class count is a parameter, so a class no sample happens to carry raises nothing — there is no inference left to be wrong about.

Members

Member	What it does
TopKAccuracy.Score	The fraction of samples whose true class is among the k highest-scoring, or the count of them.