Two-way Boosting Partial Least Squares

Abstract

Two-way Boosting Partial Least Squares (TBPLS) regression is described. TBPLS can be seen as an extension of conventional boosting techniques in sample dimensions to variable dimensions (i.e. descriptor). The basic idea of boosting is a series of regressor or classifier obtained by a part of data. In other words, this process can be seen as a kind of resampling based of prediction error by the previous model. In the present study, this idea is extended as boosting in variable dimensions. It proceeds as follows; uninformative variables in the previous are intensively selected for the next model building. This process is repeated several times and final model is obtained as an ensemble of them. Two-way boosting is the simultaneous boosting in both sample and variable dimensions in PLS modeling. This model ensemble technique makes it possible to describe much more latent relationship not only among samples but also variables, such as descriptors. The effect of TBPLS is demonstrated with a QSAR data set.