The way it works is if you have 5 known places that a user likes, show the model 4 and see if it can guess the 5th.
OK, so it’s purely on ‘type/location’ style data.
So if it’s a rubbish restaurant, it’ll still get added to my list?
Not quite, the places can be from any category/type as these types of algorithms simply look for shared interests, which can easily be across category.
In order to assess performance you can therefore take a complete list of places (these can be a complete mixture) that a user likes, remove some at random and then ask your model if it can guess which ones are missing. Spotify have quite a nice explanation of R-precision in their recommender system competition documentation.
2 Likes