recallonly

Recallonly is a term used to describe a focus on recall as the primary objective in evaluating or training a model. Recall is the proportion of relevant instances that are retrieved, calculated as true positives divided by the sum of true positives and false negatives. In a recallonly setting, other metrics, notably precision, are not considered when selecting models or thresholds.

Applications include domains where missing a positive instance has high cost, such as medical screening, early disease detection, safety-critical anomaly detection, or fraud triage. The goal is to minimize false negatives even at the expense of more false positives.

Implementation approaches include adjusting decision thresholds to favor higher recall, using loss functions that penalize false negatives more heavily, applying class-weighted learning, or employing anomaly-detection or one-class methods that optimize recall for the positive class. In practice, recall may be optimized through threshold sweeps or by designing metrics that only track recall during model selection.

Limitations of a recallonly approach include a high rate of false positives, leading to wasted human review or resource strain. It can also yield unstable results if data distributions shift, and it neglects user experience or downstream costs captured by precision-based metrics. As a result, recall-only optimization is often paired with commentary on the trade-offs or replaced with balanced metrics like the F1 score or precision-recall curves in broader evaluations.

See also: recall, precision, F1 score, information retrieval, thresholding, class imbalance, ROC AUC.