Matrix factorization for collaborative filtering - new users and items?

Question

I've been reading about using matrix factorization for collaborative filtering, but I can't seem to find an example that deals with adding a new user or item to the system, or having the user rate a new item. In these cases, the item-user matrix and the factorization needs to be recomputed, correct? How can this perform well with a large number of users and items? Is there a way around it?

Thank you

Chavey

A couple of additional terms that might help you in your search would be "online collaborative filtering" and stochastic gradient descent. I have not used the following and it is java but you may want to check out github.com/MrChrisJohnson/CollabStream as an example of a project that might address your need. – Escuage

See here for a possible solution: #41537970 – Multilingual

Answer 1

Your question has two parts: (A) How to deal with new users and items, and (B), how to deal with new interactions (e.g. ratings, clicks, etc.).

(A) There are basically 2 different strategies for dealing with new users and items (no matter whether we use matrix factorization or something else):

1. estimating user/item features from user (demographics, surveys) or item (price, genre, textual description, categories) attributes
2. active learning: showing new items to all users interacting with the system, or certain items to new users of the system, in a way balancing individual user experience and information gain by the system.

There are many papers in the academic literature on both problems.

(B) This is really not problematic -- incremental updates to a matrix factorization model does not have high computational costs. See for example this paper: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.165.8010&rep=rep1&type=pdf

The MyMediaLite library (disclaimer: I am the main author) supports incremental updates for several matrix factorization methods: http://ismll.de/mymedialite

Nomadic

If you use a factorization algorithm such as incremental svd and hence "complete" the user x item matrix and a new customer arises under the scenarios 1) they have some ratings or 2) they have no ratings how would you "score" them, without re-running the entire svd? Under scenario 1 could you fall back to performing an SVD (not incremental, but standard svd) on the "completed" matrix and then use a similarity measure to see which users they are closest to and use the entries in the completed matrix to make recommendations? – Vandervelde

If you don't mind, one more question since you seem to be an expert in this field: Can the incremental SVD (Simon Funk) be used for binary data (customer purchased or not) or does something else need to be used? Thanks! – Vandervelde