ON BETTER GENERALIZATION BY COMBINING 2 OR MORE MODELS - A QUANTITATIVE STRUCTURE-ACTIVITY RELATIONSHIP EXAMPLE USING NEURAL NETWORKS

An important problem in statistical modeling (regression or classification) is to build models that generalize well. We propose a method to combine two or more models each obtained using cross-validation (CV), to generate a model that is better at generalization compared to the original ones. Such combinations are especially useful when cross-validation experiments indicate many possible solutions. The basic idea of the procedure described comes from actively making use of the predictions of cross-validation experiments, not just for calculating the CV error. The method is applied and is shown to perform well when building quantitative structure-activity relationships (QSARs) for steroids using neural networks. The proposed method has a much broader applicability than QSAR or neural networks.