modelkit
Modelkit.Ranking_metricsDiscounted cumulative gain over per-row graded relevance.
Each row of relevance holds finite non-negative gains for the row's items and each row of scores holds the ranking scores; both matrices share one shape with at least two columns. Rank r receives discount 1 / log2 (r + 2), and k zeroes discounts from rank k onward. By default tied scores share the mean gain of their group times the group's summed discount, following McSherry and Najork; ignore_ties instead ranks tied items by descending column index as scikit-learn's reversed stable sort does. NDCG divides each row by its ideal DCG and scores an all-zero row as zero. Both metrics average rows by sample weight and cost O(rows * columns log columns).
val dcg :
?k:int ->
?ignore_ties:bool ->
?sample_weight:Sample_weight.t ->
relevance:Matrix.t ->
scores:Matrix.t ->
unit ->
(float, Error.t) resultval ndcg :
?k:int ->
?ignore_ties:bool ->
?sample_weight:Sample_weight.t ->
relevance:Matrix.t ->
scores:Matrix.t ->
unit ->
(float, Error.t) result