Random Forests and Homogeneous Granulation

Abstract

The work is a continuation of our research on the application of the newly discovered homogeneous granulation technique. The method gives the possibility to reduce the size of decision-making systems while maintaining their classification efficiency without the need to estimate the optimal approximation radii. The level of system approximation depends on the level of homogeneity of decision classes. That is, the tolerance of modification of objects with their preservation in a given class. Being motivated by effectiveness of our recently developed Ensemble model of Random Granular Reflections - where the homogeneous granulation technique was used to select objects for individual learning iterations - we have checked the effectiveness of the Random Forest in the context of boosting the classification on granular data. In the applied technique, an appropriate subset of attributes and objects is used in individual learning iterations. This means that training data is reduced in two ways. The results of experiments carried out on selected data from the UCI repository show reasonable efficiency on significantly reduced training systems.