Researching on Multi-net Systems Based on Stacked Generalization

Among the approaches to build a Multi-Net system, Stacked Generalization is a well-known model. The classification system is divided into two steps. Firstly, the level-0 generalizers are built using the original input data and the class label. Secondly, the level-1 generalizers networks are built using the outputs of the level-0 generalizers and the class label. Then, the model is ready for pattern recognition. We have found two important adaptations of Stacked Generalization that can be applyied to artificial neural networks. Moreover, two combination methods, Stacked and Stacked+, based on the Stacked Generalization idea were successfully introduced by our research group. In this paper, we want to empirically compare the version of the original Stacked Generalization along with other traditional methodologies to build Multi-Net systems. Moreover, we have also compared the combiners we proposed. The best results are provided by the combiners Stacked and Stacked+ when they are applied to ensembles previously trained with Simple Ensemble.