A recipe app tests its recommender the way the Precision and Recall at k lesson did. It hides every recipe each person cooked after a cut-off date, asks the recommender for each person's top 5, and checks each list against that person's hidden picks.
| Person | Top 5, in order (slot 1 is shown first) | Hidden picks |
|---|---|---|
| Ana | Pho, Dal, Ramen, Tacos, Gumbo | Ramen, Paella |
| Ben | Curry, Pesto, Bibimbap, Dal, Ramen | Curry |
| Cara | Tacos, Gumbo, Pho, Pesto, Curry | Paella, Dal |
| Dev | Gumbo, Paella, Curry, Ramen, Pho | Ramen, Paella |
| Eli | Dal, Ramen, Pesto, Tacos, Bibimbap | Bibimbap |
Precision@5 treats all five slots the same. But people mostly look at the first slot or two, so the team also scores how high up the list the first hit sits. A person's reciprocal rank is
and if their list has no hit at all. The mean reciprocal rank (MRR) is the average RR over everyone tested.
What is this recommender's MRR? Round to 3 decimal places.