Dimension selection for SVD models, with applications to relational data

Peter Hoff
Departments of Statistics and Biostatistics and
The Center for Statistics and the Social Sciences
University of Washington

Wednesday, January 25th
3:30pm
PWB 2-470
Minneapolis Campus

Abstract:

Matrix representation techniques have a long history in the analysis of multivariate data, including relational data in which observations are on pairs of individuals or units. In particular, the singular value decomposition of a matrix allows one to represent the relationship between two units as the inner product of a pair of latent characteristic vectors. One outstanding issue in the use of such models has been the determination of the dimension of this latent space of characteristics. In this talk I show how Bayesian methods can be used to select an appropriate dimension, and how Bayesian model averaging over the dimension can improve upon the predictive power of these models. I illustrate this with a small simulation study and two example analyses: one on conflicts among nations
during 1990-2000, the other on protein-protein interaction networks.

A social tea will be held at 3:00 P.M. in A434 Mayo. All are Welcome.
For more details contact 612-624-4655 or see http://www.biostat.umn.edu/seminar_academic.html