Authors: Geppert, Leo N.
Ickstadt, Katja
Munteanu, Alexander
Sohler, Christian
Title: Random projections for Bayesian regression
Language (ISO): en
Abstract: This article introduces random projections applied as a data reduction technique for Bayesian regression analysis. We show sufficient conditions under which the entire d -dimensional distribution is preserved under random projections by reducing the number of data points from n to k element of O(poly(d/epsilon)) in the case n >> d . Under mild assumptions, we prove that evaluating a Gaussian likelihood function based on the projected data instead of the original data yields a (1+ O(epsilon))-approximation in the l_2-Wasserstein distance. Our main result states that the posterior distribution of a Bayesian linear regression is approximated up to a small error depending on only an epsilon-fraction of its defining parameters when using either improper non-informative priors or arbitrary Gaussian priors. Our empirical evaluations involve different simulated settings of Bayesian linear regression. Our experiments underline that the proposed method is able to recover the regression model while considerably reducing the total run-time.
Issue Date: 2014-04
Appears in Collections:Sonderforschungsbereich (SFB) 876

Files in This Item:
File Description SizeFormat 
geppert_etal_2014a.pdfDNB982.28 kBAdobe PDFView/Open

This item is protected by original copyright

If no CC-License is given, pleas contact the the creator, if you want to use thre resource other than only read it.