A Robust Strategy for Combining Several Classifiers for Small Samples and Heterogeneous Predictors

Faced to safety constraints, one cannot rely on a single prediction method, especially when the sample size is low. Stacking introduced by Wolpert (1992) and Breiman (1996) is a successful way of combining several models. We modify the usual stacking methodology when the response is binary and predictions highly correlated, by combining predictions with PLS-Discriminant Analysis instead of ordinary least squares. A strategy based on repeated split samples is then developed to select relevant variables and ensure the robustness of the final model. This method is applied to the prediction of hazard of 165 chemicals, based upon 35 in vitro and in silico characteristics.