Reducing the dimensionality of dissimilarity space embedding graph kernels

Graphs are a convenient representation formalism for structured objects, but they suffer from the fact that only a few algorithms for graph classification and clustering exist. In this paper a new approach to graph classification by dissimilarity space embedding is proposed. This approach, which is in fact a new graph kernel, allows us to apply advanced classification tools while retaining the high representational power of graphs. The basic idea of the proposed graph kernel is to view the edit distances of a given graph g to a set of training graphs as a vectorial description of g. Once a graph has been transformed into a vector, different dimensionality reduction algorithms are applied such that redundancies are eliminated. To this reduced vectorial data representation any pattern classification algorithms available for feature vectors can be applied. Through various experiments it is shown that the proposed dissimilarity space embedding graph kernel outperforms conventional classification algorithms applied in the original graph domain.