Automatic Classification of Graphs by Symbolic Histograms

Abstract

An automatic classification system coping with graph patterns with node and edge labels belonging to continuous vector spaces is proposed. An algorithm based on inexact matching techniques is used to discover recurrent subgraphs in the original patterns, the synthesized prototypes of which are called symbols. Each original graph is then represented by a vector signature describing it in terms of the presence of symbol instances found in it. This signature is called symbolic histogram. A genetic algorithm is employed for the automatic selection of the relevant symbols, while a K-nn classifier is used as the core inductive inference engine. Performance tests have been carried out using algorithmically generated synthetic data sets.