Kernel k-Means Clustering Applied to Vector Space Embeddings of Graphs

In the present paper a novel approach to clustering objects given in terms of graphs is introduced. The proposed method is based on an embedding procedure that maps graphs to an n-dimensional real vector space. The basic idea is to view the edit distance of an input graph gto a number of prototype graphs as a vectorial description of g. Based on the embedded graphs, kernel k-means clustering is applied. In several experiments conducted on different graph data sets we demonstrate the robustness and flexibility of our novel graph clustering approach and compare it with a standard clustering procedure directly applied in the domain of graphs.