It would honestly be so nice to just rate a bunch of restaurants (or anything, music would be really nice) and have it spit out recommendations like “people with really similar taste to yours also like x,y,z”
That project never really reached its full potential for other reasons, but it turns out that even with a lot of data, it's hard to make big gains in ranking over just using the aggregate of all users. The intersection of two people's tastes is sort of surprisingly not very helpful.
It didn't really translate well to visiting Japan, but does great when in a new metro and not having any personal recommendations to go off.
The cross city problem is one of the things I've looked at a lot (I actually wrote a design doc for it at a different company). Even with really good coverage in New York/San Francisco/London/Tokyo due to our teams being based there, it was hard to get significant personalized ranking improvement. Maybe it's better now that they have a lot more data.
"Restaurants liked by people who like x and y and z"