How do recommender systems incorporate user characteristics? I'm new to recommender systems, and I've been reading about how user-based collaborative filtering can group similar users together and (for example) use their ratings to suggest movies to other similar users. However, I'm having trouble understanding exactly how a recommender system incorporates user characteristics into its modeling.
In the typical modeling framework (eg. linear models, decision trees, etc.) the analyst would explicitly define the variables to be used in the modeling process, thus allowing them to explicitly define the user features that go into the model. However, in a recommender system, the model is simply created from a matrix that contains only three data points (germane to the movie rating example I've been using): user ID, movie ID, and movie rating.
The matrix doesn't have any room for user characteristics like demographics or movie preferences, so how exactly can a recommender system claim to incorporate these characteristics in its modeling? Are they somehow incorporated implicitly through some kind of ordering of the matrix?
Any thoughts or resources in understanding this would be greatly appreciated.
 A: I am unaware of movie "recommender systems claiming to incorporate demographics" or other user data (except the movie scores). Yes, socio-demographic data is reflected in the scores (e.g. some films being preferred by females and others by young men), and while this might be interesting on a global level (e.g for sociologists or ethnologist observing what types of films are liked in different cultural regions or by females) and film sites may display such stats, deductions in the other direction are not possible/valid (e.g. just because a user is a woman doesn't tell you how much she is going to like a particular film because the differences within the group "women" are too important).
The movie recommending systems I know just compare individual votes, not taking into account any other user data. If you are looking for an example, in this question you will find an explanation of how one particular movie recommending system works: How to correlate movie taste?
A: I'm working on a somewhat similar problem, but more complicated because it is a social network service, and the "items" to be recommended are other users, and what you are trying to get at is something like a 'hybrid' approach, mixing in elements of collaborative filtering with content-based recommendations. I have found a few research papers that talk about this, and my intuition is that one might use both content-based and collaborative filtering to find a consensus recommendation, or use one to narrow down or sanity check the output of the other, or as I am thinking of doing, use users similar in their content (personal data) to identify users with similar preferences, and then designate those as the best collaborators whose votes should be used to recommend other users. SNS recommendation systems are all pretty proprietary though, that's probably why I can't find a good OkCupid or Tinder blog post on something like this. One paper that seems promising is this one.
