Generalised median of graph correspondences

摘要

A graph correspondence is defined as a function that maps the elements of two attributed graphs. Due to the increasing availability of methods to perform graph matching, numerous graph correspondences can be deducted for a pair of attributed graphs. To obtain a representative prototype for a set of data structures, the concept of the median has been largely employed, as it has proven to deliver a robust sample. Nonetheless, the calculation of the exact (or generalised) median is known to be an NP-complete problem for most domains. In this paper, we present a method based on an optimisation function to calculate the generalised median graph correspondence. This method makes use of the Correspondence Edit Distance, which is a metric that considers the attributes and the local structures of the graphs to obtain more interesting and meaningful results. Experimental validation shows that this approach is capable of obtaining the generalised median in a comparable runtime with respect to state-of-the-art methods on artificial data, while maintaining the success rate for a real-application case. (C) 2019 Elsevier B.V. All rights reserved.