A comparison between structural and embedding methods for graph classification

Abstract

Structural pattern recognition is a well-know research field that has its birth in the early 80s. Throughout 30 years, structures such as graphs have been compared through optimization of functions that directly use attribute values on nodes and arcs. Nevertheless, in the last decade, kernel and embedding methods appeared. These new methods deduct a similarity value and a final labelling between nodes through representing graphs into a multi-dimensional space. It seems that lately kernel and embedding methods are preferred with respect to classical structural methods. However, both approaches have advantages and drawbacks. In this work, we compare structural methods to embedding and kernel methods. Results show that, with the evaluated datasets, some structural methods give slightly better performance and therefore, it is still early to discard classical structural methods for graph pattern recognition. © 2012 Springer-Verlag Berlin Heidelberg.