Structural graph matching using the EM algorithm and singular value decomposition

This paper describes an efficient algorithm for inexact graph matching. The method is purely structural, that is to say, it uses only the edge or conn ectivity structure of the graph and does not draw on node or edge attribute s. We make two contributions. Commencing from: a probability distribution f or matching errors, we show how the problem of graph matching can be posed as maximum-likelihood estimation using the apparatus of the EM algorithm. O ur second contribution is to cast the recovery of correspondence matches, b etween the graph nodes in a matrix framework. This allows us to efficiently recover correspondence matches, using singular value decomposition. We exp eriment with the method on both real-world and synthetic data. Here, we dem onstrate that the method offers comparable performance to more computationa lly demanding methods.