## Academic DefenseOn Thursday May 6 2010 at 14h00 in Room 1 - IRMA Tower
Summary We present a general procedure for nonparametric multivariate regression smoothers that outperforms existing procedures such as MARS, additive models, projection pursuit or $L_2$ additive boosting on both real and simulated datasets. In multivariate nonparametric analysis, sparseness of the covariates also called curse of dimensionality, forces one to use large smoothing parameters. This leads to biased smoother. We still propose to use classical nonparametric linear smoother, such as thin plate splines or kernel smoothers, but instead of focusing on optimally selecting the smoothing parameter, we fix it to some reasonably large value to ensure an over-smoothing of the data. The resulting (base) smoother has a small variance but a substantial bias. Afterward, we propose to iteratively correct the biased initial estimator by an estimate of the bias obtained by smoothing the residuals. In univariate settings, we relate our procedure to $L_2$-Boosting. Rules for selecting the optimal number of iterations are also proposed and, based on empirical evidence, we propose one stopping rule. In the regression framework, when the unknown regression function m belongs to the Sobolev space H(n) of order n, we show that using a thin plate splines base smoother and the proposed stopping rule leads to an estimate ^m which converge to the unknown function m. Moreover, our procedure is adaptive with respect to the unknown order n and converge at the minimax rate. We apply our method to both simulated and real data and show that our method compares favourably with existing procedures such as MARS, additive models, L2 boosting or projection pursuit, regression trees, with improvement on mean squared error up to 30%. A R package is available at http://cran.r-project.org/web/packages/ibr/index.html |