🎓 Learning corner

Ranking and retrieval metrics

Use these metrics when a system returns an ordered list: search results, similar documents, recommendations, or ranked answers.

Background

Search, recommendations, semantic retrieval, and question answering often return a ranked list. For those tasks, we care not only whether the right item appears, but how far down the list users must look.

Question Correct result appears at... Recall@1 credit MRR@10 credit
Q1 Rank 1 1 1.00
Q2 Rank 2 0 0.50
Q3 Rank 12 0 0.00
Q4 Rank 1 1 1.00
Q5 Rank 5 0 0.20

Recall@1

Only checks the first result. Here, 2 out of 5 top results are correct, so Recall@1 is 40%.

MRR@10

Gives partial credit when the first correct result appears within the top 10. Here, the average is 0.54.

⚔️ Spearman

Spearman is used when the exact score matters less than the order. It asks whether two ranked lists place the same items in a similar order.

Spearman example comparing two movie rankings with lines connecting corresponding movies.
Spearman value Plain meaning
+1 The two rankings perfectly agree.
Around 0 There is no clear relationship between the rankings.
-1 The two rankings are exactly reversed.

🥇 Mean Reciprocal Rank

Mean Reciprocal Rank, or MRR, gives more credit when the first correct result appears higher in the list. Rank 1 gets full credit. Rank 2 gets half credit. Rank 3 gets one-third credit.

Mean reciprocal rank chart showing that lower ranked correct answers receive less credit.

MRR@10 uses the same idea but only checks the top 10 results. If the first correct result appears below rank 10, it receives no credit for that query.