Résumé
This work deals with mutliblocks data analysis and especially with regression. This thesis presents, like a collection of methods, a large number of methods of mutli-block data analysis, bad known for the majority of them. This thesis contains three parts. The first one, with seven chapters, presents methods for linear modelling. They are classified by kind of analysis (simultaneous or regression) and by applicability domain (multi-block or multiway). We propose several new PLS-like one's, stepwise constrained optimization based, dedied to the simultaneous analysis of one (OMCIA) or more sets (GOMCIA1,2,3) of matrices and to model a set of matrices with another one (OMCIA-PLS, GOMCIA-PLS1,2,3, MUBRE), OMCIA allowing to obtain better trajectories in the STATIS method. Moreover, we propose a generalization of the STATIS method to two multiblock tables (DO-ACT). A chapter is dedied to the practice of the multiblock methods, in which propose some criteria to choose the important predictive blocks in regression multi-block models. <br /><br />In the second part, we deal with the non-linear modelling. We start to give a review of non-linear PLS methods for two matrices and then propose two non-linear multiblock regression methods (MUBRAS), searching for the optimal spline tranformations of the predictors. To end, in the third part, we present several applications of our methods, programmed in S-Plus\textsuperscript{\textregistered} language, on real data from chemometrics, sensory analysis, ecology .....