Iterative Bipartite Graph Edit Distance Approximation

Abstract

One of the major tasks in many applications in the field of document analysis is the computation of dissimilarities between two or more objects from a given problem domain. Hence, employing graphs as representation formalism evokes the need for powerful, fast and flexible graph based dissimilarity models. Graph edit distance is powerful and applicable to any kind of graphs but suffers from its high computational complexity. Recently, however, a novel framework for graph edit distance approximation has been introduced. While the run time of this novel procedure is very convincing, the precision of the approximated graph distances is dissatisfying in some cases. The present paper introduces a generalized version of the existing approximation framework using an iterative bipartite procedure. With empirical investigations on three real world data sets we show that our extension substantially improves the accuracy of the approximations while the run time is increased only linearly with the number of additional iterations.