Metrics

In this article we provide a list of predefined metrics that can be used during model training and testing.

Metrics can be passed by their name via metrics parameter in TrainingParams and TestingParams. To read more, visit Fine-tuning training parameters and Customizing testing metrics.

Metric name	Task	Description
MultipleTargetsRecall	Multiclass Classification	A fraction of correct class predictions. The result is one number.
MultipleTargetsRecallPerClass	Multiclass Classification	The number of correctly predicted positive instances divided by the total number of actual positive instances for a given class. This yields one value (recall) per class.
PrecisionAtK	Recommendations	The fraction of the top K predicted items that are relevant, computed for each instance and averaged over the dataset.
MeanAveragePrecisionAtK	Recommendations	The average precision at each position of relevant items within the top K results, giving higher scores when relevant items appear earlier in the ranking, and then averages this over all instances.
HitRateAtK	Recommendations	The fraction of cases where the top K predictions contain at least one relevant item.
MeanReciprocalRank	Recommendations	A ranking metric that measures how early the first relevant item appears in a list of recommendations, defined as the average reciprocal rank of the first relevant item across all entities (e.g. customers), with entities that have no relevant item contributing zero.
NDCGAtK	Recommendations	The measure of how well the top K results are ordered, rewarding highly relevant items appearing earlier in the list and normalizing the score against the ideal ranking so values lie between 0 and 1.