We propose a feature selection method for density estimation with quadratic loss. This method relies on the study of unidimensional approximation models and on the definition of confidence regions for the density thanks to these models. It is quite general and includes cases of interest like detection of relevant wavelets coefficients or selection of support vectors in SVM. In the general case, we prove that every selected feature actually improves the performance of the estimator. In the case where features are defined by wavelets, we prove that this method is adaptative near minimax (up to a log term) in some Besov spaces. We end the paper by simulations indicating that it must be possible to extend the adaptation result to other features.
Keywords: density estimation, support vector machines, kernel algorithms, thresholding methods, wavelets
@article{PS_2008__12__438_0,
author = {Alquier, Pierre},
title = {Density estimation with quadratic loss : a confidence intervals method},
journal = {ESAIM: Probability and Statistics},
pages = {438--463},
year = {2008},
publisher = {EDP-Sciences},
volume = {12},
doi = {10.1051/ps:2007050},
mrnumber = {2437718},
language = {en},
url = {http://geodesic.mathdoc.fr/articles/10.1051/ps:2007050/}
}
TY - JOUR AU - Alquier, Pierre TI - Density estimation with quadratic loss : a confidence intervals method JO - ESAIM: Probability and Statistics PY - 2008 SP - 438 EP - 463 VL - 12 PB - EDP-Sciences UR - http://geodesic.mathdoc.fr/articles/10.1051/ps:2007050/ DO - 10.1051/ps:2007050 LA - en ID - PS_2008__12__438_0 ER -
Alquier, Pierre. Density estimation with quadratic loss : a confidence intervals method. ESAIM: Probability and Statistics, Tome 12 (2008), pp. 438-463. doi: 10.1051/ps:2007050
Cité par Sources :