[Statlist] Two Statistics Seminars at Bern

Lutz Duembgen |utz@duembgen @end|ng |rom @t@t@un|be@ch
Mon Jul 12 10:59:28 CEST 2004


From: Lutz Duembgen <duembgen using stat.unibe.ch>

==========================================================

Institute of Mathematical Statistics and Actuarial Science
University of Bern
Sidlerstrasse 5
CH-3012 Bern

We are pleased to announce the following two seminars:

----------------------------------------------------------

Tuesday, August 3, 2004
17:00 - 18:00, B78

Prof. Wolfgang Polonik
(University of California at Davis)

Mode Hunting in Multi-Dimensions: Data Analytic Tools
with Measures of Significance

Abstract:
In this talk we propose a novel nonprametric method for finding modes in 
multi-dimensional data sets without specifying their total number. 
Usually for such methods there exists a trade-off between data analytic 
considerations and theoretical justification. Our method addresses both 
of these issues. It is practically feasible even in moderate dimensions, 
and we also provide information on significance of the findings. The 
latter is accomplished via testing for the presence of antimodes. 
Critical values for these tests are based on large sample 
approximations. The proposed method is complemented by diagnostic plots. 
We illustrate the method by real data applications.

----------------------------------------------------------

Friday, August 13, 2004
17:00 - 18:00, B78

Prof. Guenther Walther
(Stanford University)

Stagewise algorithms for regression and the LASSO

Abstract:
The Lasso (Tibshirani 1996) is a method for regularizing least squares 
regression with L1 constraints, which leads to sparse solutions. The 
entire sequence of Lasso solutions can be computed efficiently with the 
LAR (least angle regression, Efron et al. 2003) algorithm, which also 
provides a conceptual link between Lasso and Forward Stagewise 
regression. The latter is an important component in adaptive regression 
procedures such as boosting, and hence this link helps us to understand 
how boosting works. We give a sequential criterion involving a minimum 
L1 arc-length penalty that is optimized by Forward Stagewise regression. 
We also characterize problems for which the coefficient curves for Lasso 
are monotone as a function of the L1 norm, which implies that all three 
procedures (LAR, Lasso and Forward Stagewise) coincide. This is joint 
work with Trevor Hastie, Jonathan Taylor and Rob Tibshirani.




More information about the Statlist mailing list