In model evaluation, what does recall refer to?

Recall measures the ability of a model to correctly identify all relevant instances within a dataset, specifically focusing on the true positives. It is defined as the ratio of true positives to the total actual positives, providing insight into how well the model is performing in capturing the positive cases. In contexts such as medical diagnosis or fraud detection, where it is crucial to identify as many positive cases as possible, having a high recall indicates that the model is efficient at minimizing false negatives.

In contrast, the other options do not accurately describe recall. The total number of predictions made refers to the model's workload but does not reflect its accuracy or ability to detect positives. Total accuracy encompasses a broader measure, including true positives, true negatives, false positives, and false negatives, rather than focusing on just the relevant positives. Finally, the number of true negatives pertains to the correct identification of negatives, which is not what recall is concerned with, as it is specifically focused on positive identifications. Thus, the characterization of recall as the ratio of true positives to actual positives provides a clear and precise understanding of this critical metric in model evaluation.