Graph Kernels: Crossing Information from Different Patterns Using Graph Edit Distance
Graph kernels allow to define metrics on graph space and constitute thus an efficient tool to combine advantages of structural and statistical pattern recognition fields. Within the chemoinformatics framework, kernels are usually defined by comparing the number of occurences of patterns extracted from two different graphs. Such a graph kernel construction scheme neglects the fact that similar but not identical patterns may lead to close properties. We propose in this paper to overcome this drawback by defining our kernel as a weighted sum of comparisons between all couples of patterns. In addition, we propose an efficient computation of the optimal edit distance on a limited set of finite trees. This extension has been tested on two chemoinformatics problems.
KeywordsLinear Pattern Edit Operation Statistical Pattern Recognition Graph Kernel Graph Edit Distance
- 1.Brun, L., Gaüzére, B., Fourey, S.: Relationships between graph edit distance and maximal common unlabeled subgraph. Technical report, CNRS UMR 6072 GREYC (2012), http://hal.archives-ouvertes.fr/hal-00714879
- 4.Gaüzère, B., Brun, L., Villemin, D.: Two new graph kernels and applications to chemoinformatics. Pattern Recognition Lett. (in Press, 2012)Google Scholar
- 5.Haussler, D.: Convolution kernels on discrete structures. Technical report, Dept. of Computer Science, University of California at Santa Cruz (1999)Google Scholar
- 6.Kashima, H., Tsuda, K., Inokuchi, A.: Kernels for graphs, ch. 7, pp. 155–170. MIT Press (2004)Google Scholar
- 8.Neuhaus, M., Bunke, H.: Bridging the gap between graph edit distance and kernel machines. World Scientific Pub. Co. Inc. (2007)Google Scholar