Craig Boutilier
Department of Computer Science
University of Toronto
Toronto, ON M5S 3H5
email: cebly@cs.toronto.edu
Richard S. Zemel
Department of Computer Science
University of Toronto
Toronto, ON M5S 3H5
email: zemel@cs.toronto.edu
Abstract
Collaborative filtering allows the preferences of multiple users
to be pooled in a principled way in order to make recommendations
about products, services or information unseen by a specific user.
We consider here the problem of online and interactive collaborative filtering:
given the current ratings and recommendations associated with a user,
what queries (new ratings) would most improve the quality of the
recommendations made? This can be cast in a straightforward fashion in
terms of \emph{expected value of information}; but the online computational
cost of computing optimal queries is prohibitive.
We show how offline pre-computation of bounds on value of information,
and of prototypes in query space, can be used to dramatically reduce
the required online computation. The framework we develop is quite general,
but we derive detailed bounds for the multiple-cause vector quantization
model, and empirically demonstrate the value of our active approach
using this model.
To appear, AI-Stats 2003
Return to List of Papers