Abstract:
|
High-dimensional problems abound in applied statistics. Practitioners often require methods to reduce the dimension of a given data set while retaining the salient information contained therein in order to aid decision-making processes such as assessing variable importance, providing visualization, or performing variable selection. Dimension reduction methods seek a projection of data into a lower-dimensional subspace that preserves or exploits some desired underlying structure. Unfortunately, a number of dimension reduction techniques assume a single source of data, and many further assume identically distributed data, perhaps conditioned on some variable of interest. We propose a framework to both incorporate knowledge of response variables of interest into dimension reduction techniques as well as take into account nuisance variables that affect the homogeneity of the data distribution. The framework extends to estimation of sparse projections, allowing for variable selection in addition to removing heterogeneity. We apply our methods to a variety of high-dimensional data analysis problems in image analysis.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.