摘要：This paper presents the Partial Least-Squares regression (PLS) in the framework of the boosting methods with L2 loss. First, the ordinary PLS regression already belongs to that family by considering the latent variables or principal components as base learners producing robust linear models that overcome the problems of the scarcity of the observations as well as the multi-collinearity of the predictor variables. Most of all, the use B-splines and their tensor products to construct the base learner, typically provides PLS with L2-boosts leading to non-linear additive models that capture main effects as well as relevant interactions.
The performances of the different PLS boosts in both regression and classification are shown on three exemples.
