Recall@k measures the fraction of relevant items in a user's test set that appear among the top-k recommendations. If s relevant items are retrieved from M relevant test items, recall@k = s/M. Unlike precision, recall asks whether the recommendation list captured the items that mattered, not how many irrelevant items it also contained. section 4.6.1
Use the equation to relate the number of captured relevant items to the user's complete relevant test set. section 4.6.1
Temporal test-set example
A user consumed [4, 7, 25, 53, 64, 5]. After temporal splitting, [4, 7, 25, 53] belong to training and [64, 5] form the test set. If both 64 and 5 appear in the top-10 recommendations, recall@10 is 2/2 = 1.0. Increasing k generally raises recall because more items can capture the relevant set, while precision usually falls as more irrelevant items enter the list. section 4.6.1
Recall is especially important in retrieval, the wide first stage of a multi-stage recommender. A candidate pool succeeds when it contains relevant items for later scoring, even if most candidates are mediocre. Measure recall at pool sizes such as 50, 100, 200, and 500, then choose the smallest pool that captures most relevant items near the elbow of the recall curve. Ranking evaluates position; retrieval primarily evaluates presence. section 4.6.5


































