Extending Flow Graphs for Handling Continuous−Valued Attributes

Abstract

This paper describes a research work that focuses on a suitable data structure, capable of summarizing a given supervised training data set into a weighted and labeled digraph. The data structure, named flow graph (FG), has been proposed not only for representing a set of supervised training data aiming at its analysis but, also, for supporting the extraction of decision rules, aiming at a classifier. The work described in this paper extends the original FG, suitable for discrete data, for dealing with continuous data. The extension is implemented as a discretization process, carried out as a pre-processing step previously to learning, in a two-step hybrid approach named HFG (Hybrid FG). The results of the conducted experiments were analyzed with focus on both, the induced graph-based structure and the performance of the associated set of rules extracted from the structure. Results obtained using the J48 are also presented, for comparison purposes.