Isomorphism Testing via Polynomial-Time Graph Extensions

摘要

This paper deals with algorithms for detecting graph isomorphism (GI) properties. The GI literature consists of numerous research directions, from highly theoretical studies (e.g. defining the GI complexity class) to very practical applications (pattern recognition, image processing). We first present the context of our work and provide a brief overview of various algorithms developed in such disparate contexts. Compared to well-known NP-complete problems, GI is only rarely tackled with general-purpose combinatorial optimization techniques; however, classical search algorithms are commonly applied to graph matching (GM). We show that, by specifically focusing on exploiting isomorphism properties, classical GM heuristics can become very useful for GI. We introduce a polynomial graph extension procedure that provides a graph coloring (labeling) capable of rapidly guiding a simple-but-effective heuristic toward the solution. The resulting algorithm (GI-Ext) is quite simple, very fast and practical: it solves GI within a time in the region of O(|V|3) for numerous graph classes, including difficult (dense and regular) graphs with up to 20.000 vertices and 200.000.000 edges. GI-Ext can compete with recent state-of-the-art GI algorithms based on well-established GI techniques (e.g. canonical labeling) refined over the last three decades. In addition, GI-Ext also solves certain GM problems, e.g. it detects important isomorphic structures induced in non-isomorphic graphs.