Kernel Logistic PLS” (KL-PLS), a new tool for classification with performances similar to the most powerful statistical methods is described in this paper. KL-PLS is based on the principles of PLS generalized regression and learning via kernel. The successions of simple regressions, simple logistic regression and multiple logistic regressions on a small number of uncorrelated variables that are computed within KL-PLS algorithm are convenient for the management of very high dimensional data. The algorithm was applied to a variety of benchmark data sets for classification and in all cases, KL-PLS demonstrates its competitivenesswith other state-of-art classification method. Furthermore, leaning on statistical tests related to the logistic regression, KL-PLS allows the systematic detection of data points close to “support vectors” of SVM and thus reduces the computationalcharges of the SVM training algorithm without significant loss of accuracy.