Abstract:

Research on reasoning about user preferences in artificial
intelligence attempts both to connect between modeling principles
coming from philosophy, psychology, and economics, and to develop
computational tools grounded in these principles. The goal is to
provide users with effective, computationally efficient
decision-support tools that will assist users in their everyday
activities (e.g., choosing an attractive vacation package.) Extensive
research in this direction significantly reduced the gap between
theory and practice of decision theory. However, the gap is still
there. The main problem to date is in the absence of a universal
solution for the problem that is both efficient and effective for all
sets of choices and all types of preference information.

In this talk we present the first approach that gets close to these
"universality" goals for the task of reasoning about ordinal user
preferences. Our proposal is based on connecting ideas and techniques
from the areas of  knowledge representation, philosophical logic,
convex optimization, and machine learning.  The mathematical framework
we propose is based on a novel high-dimensional structure for
preference decomposition and computational techniques adopted from
Support Vector Machines (SVMs).

Joint work with Thorsten Joachims (Cornell)