seminário

[Francisco Cribari <cribari@de.ufpe.br>]

   PROXIMO SEMINARIO DO MESTRADO EM ESTATISTICA DA UFPE
   ----------------------------------------------------
   [ conjunto c/ doutorado em matematica computacional ]


   Bolstered Error Estimation

   Ulisses Braga-Neto
   University of Texas MD Anderson Cancer Center and Texas A&M University

   RESUMO:
   Functional Genomics applications depend fundamentally on estimating
   the probability of error of designed classifiers. Error estimation
   is well-known to be problematic with the small-sample data that are
   typically associated with microarray-based studies. We present a novel
   error estimation technique that is suitable for small-sample data.
   It displays low variance and generally low bias as well, and it leads
   to very fast error estimators. This technique is based on "bolstering"
   the original empirical distribution of the data. We discuss
   small-sample classification and feature ranking experiments, using
   synthetic and real patient data, which demonstrate the performance
   of bolstered error estimation. In these experiments, bolstered error
   estimators vastly improved on resubstitution and cross-validation,
   and they were competitive with, and in many occasions superior to,
   bootstrap error estimators, while being tens to hundreds of times
   faster.


   Data: 2 de junho de 2004 (quarta-feira)
   Horario: 16:00 horas
   Local: Sala 11 (sala de aulas da PG)

--
Francisco Cribari-Neto               voice: +55-81-21267425
Departamento de Estatistica          fax:   +55-81-21268422
Universidade Federal de Pernambuco   e-mail: cribari@de.ufpe.br
Recife/PE, 50740-540, Brazil         http://www.de.ufpe.br/~cribari

       All roads lead to Rome.          --Roman proverb
       All roads do not lead to Rome.   --Slovenian proverb