What is recall? How many actual targets the system finds
Explain recall, missed positives and reference scope in classification and retrieval.
Understand the concept
Recall is TP/(TP+FN): correctly found positives among all actual positives. Unlike precision, it measures misses and requires the true target set, not only returned results. High recall can coexist with false positives and is not standalone answer accuracy.
Missed real targets belong in the denominator
Google and scikit-learn define recall through actual positives and misses. Finding four of five spam messages gives four divided by five. Extra ordinary messages affect precision, not target count. Retrieval needs relevance labels; one hit is not complete recall. Incomplete references only support labeled-scope results, not all possible material.
Keep complete references to locate omissions
Label answering passages in public documents, query with fixed settings and inspect candidate omissions. State result cutoffs and relevance criteria and separate retrieval from generation quality. Classification also needs missed positives, not only returned predictions. No index or classifier was run. Update references when the relevant set changes.
More candidates do not automatically improve the final task
More results can improve recall while adding noise and latency. Inspect precision and final quality rather than selecting everything for a high ratio. Include labeling and omission review. Preserve queries, versions and cutoffs for reruns. Recall is neither per-answer confidence nor guaranteed business improvement.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.