Coverage in Selective Prediction Explainer
Coverage in selective prediction is the proportion of inputs for which the system chooses to answer. If a model predicts on 80 of 100 cases and defers 20, its coverage is 80 percent.
Lowering coverage often lets a model keep only higher-confidence cases, reducing error among accepted predictions. The paired quantity, selective risk, measures loss or error on the cases the system actually answered.
A coverage target has operational consequences. Lower coverage increases work for humans or fallback systems and may delay service. It can also affect groups differently if uncertainty scores are less reliable for some populations.
Teams evaluate a risk-coverage curve rather than choosing a threshold from confidence alone. The final operating point should reflect error costs, reviewer capacity, fairness, urgency, and how well the fallback works in practice.
Coverage is the share of cases on which a selective prediction system chooses to provide an answer instead of abstaining or deferring.
Key terms include coverage, accepted case, abstained case, selective risk, risk-coverage curve, operating point, workload and fallback.
The denominator includes all eligible cases presented to the selective system.
The numerator includes only those for which it returns a prediction.
Lower coverage can improve answered-case reliability while expanding the review queue.
Threshold selection should include service levels and subgroup effects.
Explore more "Explainers"
Discover additional explainers across politics, science, business, technology, and other fields. Each explainer breaks down a complex idea into clear, everyday language—helping you better understand how major concepts, systems, and debates shape the world around us.