Machine learning with seriated graphs

Abstract

The aim in this paper is to show how the problem of learning the class-structure and modes of structural variation in sets of graphs can be solved by converting the graphs to strings. We commence by showing how the problem of converting graphs to strings, or seriation, can be solved using semi-definite programming (SDP). This is a convex optimisation procedure that has recently found widespread use in computer vision for problems including image segmentation and relaxation labelling. We detail the representation needed to cast the graph-seriation problem in a matrix setting so that it can be solved using SDP. We show how the strings delivered by our method can be used for graph-clustering and the construction of graph eigenspaces. © Springer-Verlag Berlin Heidelberg 2005.