[Statlist] Seminar fuer Statistik

Christina Kuenzli kuenz|| @end|ng |rom @t@t@m@th@ethz@ch
Wed May 19 11:54:22 CEST 2004


                       ETH and University of Zurich 
      Proff. A.D. Barbour -- P. Buehlmann -- F. Hampel -- H.R. Kuensch 

**************************************************************************
           We are pleased to announce the following seminar
**************************************************************************

Friday, May 28, 2004, 15.15, LEO C 15

     Andreas Buja, The Wharton School, University of Pennsylvania

     Degrees of Boosting -
     A Study of Loss Functions for Classification and Class
     Probability Estimation 

We examine a vast class of loss functions for binary
classification and class probability estimation.  The elements of this
class are known in subjective probability as "proper scoring rules". They
comprise  all common loss functions, including Kullback-Leibler 
loss, squared error loss, boosting loss derived from exponential loss, and
all cost-weighted misclassification losses.

We show that all loss functions can be interpreted as mixtures of
cost-weighted misclassification losses. The implications are:

1) We obtain immediate insight into the nature of specific loss functions.
2) We can focus loss functions on specific ranges of class probabilities,
   with good success in linear logistic regression as well as tree-based
   classification.
3) We can derive boosting-like re-weighting algorithms from the Iteratively
   Reweighted Least Squares implementation of Fisher Scoring,
4) We can extend the mixture decomposition to Bregman distances and
   use it to derive inequality bounds for asymptotic theory.

Joint work with W. Stuetzle and Y. Shen

**************************************************************************

(LEO (Leonhardstrasse 27, 8006 Zurich) is close to the main building,
across the hill-side station of the 'Polybahn') 
Overview maps of ETH : http://www.ethz.ch/search/orientation_en.asp

Further information: Christina Kuenzli, Statistics Seminar of ETH Zurich 
**************************************************************************

Everybody is kindly invited
 
Eidgenoessische Technische Hochschule Zuerich
Swiss Federal Insitute of Technology Zurich
________________________________________________________
Christina Kuenzli            <kuenzli using stat.math.ethz.ch>
Seminar fuer Statistik      
Leonhardstr. 27,  LEO D11          phone: +41 1 632 3438         
ETH-Zentrum,                       fax  : +41 1 632 1228 
CH-8092 Zurich, Switzerland        http://stat.ethz.ch/~




More information about the Statlist mailing list