A Study in Granular Computing: Homogenous Granulation

This paper is presenting a new method of decision systems granulation in the family of methods inspired by Polkowski standard granulation algorithm. The new method is called homogenous granulation. The idea is to create the granules around each training object separately by selecting smallest r-indiscernibility ratio, based on which granule consists of group of objects with the same class. This is natural idea, where the indiscernibility level is extended until indiscernibility class contains uniform group of objects. After granulation process we have used random choice for covering of universe of objects and majority voting to create granular reflections of selected granules. The main advantage of this method is lack of necessity to estimate optimal granulation radius. We have performed experiments on data from UCI repository using 5 times cross validation 5 model. First results of homogenous granulation, in the terms of classification accuracy, are comparable with the ones of already presented algorithms with significant reduction of training data size after granulation.