Compressed Regression
Shuheng Zhou, John Lafferty, Larry Wasserman Carnegie Mellon University

M30

We study high dimensional regression where the data are compressed by a random linear transformation. Motivation: scalability and privacy
 m     ×n                        +                   

m

=
random matrix

X

×1

compressed response

uncompressed data

n×p

unknown (sparse)

p×1

noise

n×1

Theoretical results:

· Bounds on number of projections that allows accurate estimation · Analysis of risk consistency · Upper bounds on information rate of compressed data