Usually, when speaking of user-based collaborative filtering, I
mentioned that the idea was great in the 90s when the internet was
already too large, but the user base of any given site was pretty
small. Well, that's not true since in a cold-start scenario you
employ specifically that method. When you have some features of all
users, you can construct a similarity matrix between them, e.g.
between new and old ones. Then you can get a new user profile as a
weighted average of her nearest neighbours in a chosen metric. Voila,
and you're are doing good-old user-based collaborative filtering, the
ancient method invented at the dawn of this research field.
