En savoir plus
Model choice
and model aggregation
For over fourty years, choosing a statistical model thanks to data
consisted in optimizing a criterion based on penalized likelihood
(H. Akaike, 1973) or penalized least squares (C. Mallows, 1973). These
methods are valid for predictive model choice (regression, classification)
and for descriptive models (clustering, mixtures). Most of their
properties are asymptotic, but a non asymptotic theory has emerged at
the end of the last century (Birgé-Massart, 1997). Instead of choosing the
best model among several candidates, model aggregation combines
different models, often linearly, allowing better predictions. Bayesian
statistics provide a useful framework for model choice and model
aggregation with Bayesian Model Averaging.
In a purely predictive context and with very few assumptions,
ensemble methods or meta-algorithms, such as boosting and random
forests, have proven their efficiency.
This volume originates from the collaboration of high-level specialists :
Christophe Biernacki (Université de Lille I), Jean-Michel Marin (Université
de Montpellier), Pascal Massart (Université de Paris-Sud), Cathy
Maugis-Rabusseau (INSA de Toulouse), Mathilde Mougeot (Université
Paris Diderot), and Nicolas Vayatis (École Normale Supérieure de Cachan)
who were all speakers at the 16th biennal workshop on advanced
statistics organized by the French Statistical Society. In this book, the
reader will find a synthesis of the methodologies' foundations and of
recent work and applications in various fields.
The French Statistical Society (SFdS) is a non-profit organization that
promotes the development of statistics, as well as a professional body for all
kinds of statisticians working in public and private sectors. Founded in
1997, SFdS is the heir of the Société de Statistique de Paris, established in
1860. SFdS is a corporate member of the International Statistical Institute and
a founding member of FENStatS - the Federation of European National
Statistical Societies.
