Metrics Taxonomy

WarpRec includes 40 GPU-accelerated metrics organized into 8 families. All metrics are implemented as PyTorch modules and support distributed evaluation.

Family	Metric	Description	Type
Accuracy	AUC	Area Under the ROC Curve.	Global
	F1@K	Harmonic mean of Precision@K and Recall@K (or custom pair).	Top-K
	GAUC	Per-user AUC averaged across all users.	Global
	HitRate@K	Fraction of users with at least one relevant item in top K.	Top-K
	LAUC@K	AUC limited to the top-K ranked items.	Top-K
	MAP@K	Mean Average Precision rewarding higher-ranked correct items.	Top-K
	MAR@K	Mean Average Recall indicating progressive retrieval quality.	Top-K
	MRR@K	Mean Reciprocal Rank of the first relevant item.	Top-K
	nDCG@K	Normalized Discounted Cumulative Gain (exponential relevance).	Top-K
	nDCGRendle2020@K	nDCG with binary relevance following Rendle et al. (2020).	Top-K
	Precision@K	Proportion of relevant items in the top K.	Top-K
	Recall@K	Proportion of relevant items successfully retrieved in top K.	Top-K
Rating	MAE	Mean Absolute Error between predicted and actual ratings.	Global
	MSE	Mean Squared Error between predicted and actual ratings.	Global
	RMSE	Root Mean Squared Error (same-scale error measure).	Global
Coverage	ItemCoverage@K	Number of unique items recommended across all users.	Top-K
	UserCoverage@K	Number of users with at least one recommended item.	Top-K
	NumRetrieved@K	Average number of items retrieved per user in top K.	Top-K
	UserCoverageAtN	Number of users with at least N items retrieved.	Top-K
Novelty	EFD@K	Expected Free Discovery (log-discounted novelty).	Top-K
	EPC@K	Expected Popularity Complement (linear novelty).	Top-K
Diversity	Gini@K	Gini Index measuring inequality of item exposure.	Top-K
	ShannonEntropy@K	Information entropy over item recommendation frequencies.	Top-K
	SRecall@K	Subtopic Recall measuring feature/category coverage.	Top-K
Bias	ACLT@K	Average Coverage of Long-Tail items in recommendations.	Top-K
	APLT@K	Average Proportion of Long-Tail items per user.	Top-K
	ARP@K	Average Recommendation Popularity of top-K items.	Top-K
	PopREO@K	Popularity-based Ranking-based Equal Opportunity.	Top-K
	PopRSP@K	Popularity-based Ranking-based Statistical Parity.	Top-K
Fairness	BiasDisparityBD@K	Relative disparity between recommendation and training bias.	Top-K
	BiasDisparityBR@K	Bias in recommendation frequency per user-item cluster pair.	Top-K
	BiasDisparityBS	Bias in training data per user-item cluster pair.	Global
	ItemMADRanking@K	Mean Absolute Deviation of discounted gain across item clusters.	Top-K
	ItemMADRating@K	Mean Absolute Deviation of average ratings across item clusters.	Top-K
	REO@K	Ranking-based Equal Opportunity across item clusters.	Top-K
	RSP@K	Ranking-based Statistical Parity across item clusters.	Top-K
	UserMADRanking@K	Mean Absolute Deviation of nDCG across user clusters.	Top-K
	UserMADRating@K	Mean Absolute Deviation of average scores across user clusters.	Top-K
Multi-objective	EucDistance@K	Euclidean distance from model performance to a Utopia Point.	Top-K
	Hypervolume@K	Volume of objective space dominated relative to a Nadir Point.	Top-K