Special Colloquium
Jie Yang
University of Chicago, Department of Statistics
Exchangeable Cluster Processes for Classification Problems
Abstract: In this talk, we introduce two families of exchangeable
cluster processes for classification problems. The first family
called the Gauss-Ewens cluster process is generated from the
Fisher discriminant model by a standard Dirichlet allocation
scheme. It permits a new unit to be assigned to a class that has
not been observed. The second family is based on the permanent
process introduced recently. In the stochastic classification
model determined by the permanent cluster process, no more than
4-5 parameters need to be estimated, regardless of the number of
classes or the dimension of the feature space. The model works
well even if a class occupies a non-convex region or disconnected
regions in the feature space. Under the permanent model, the
conditional distribution of a subsequent unit given the training
data is expressed in terms of ratios of weighted permanents. We
propose analytic approximations for the permanent ratios which is
reasonably accurate for typical classification problems.
Friday February 24, 2006 at 3:00 PM in SEO 636