A generalized Weisfeiler-Lehman graph kernel

Abstract After more than one decade, Weisfeiler-Lehman graph kernels are still among the most prevalent due to their remarkable predictive performance and time complexity. They based on a fast iterative partitioning of vertices, originally designed for deciding isomorphism with one-sided error. The retain this idea compare such labels respect equality. This binary valued comparison is, however, arguably too rigid defining suitable certain classes. To overcome limitation, we propose generalization which takes into account natural finer grade similarity between We show that proposed can be calculated efficiently by means Wasserstein distance vectors representing labels. other facts give rise choice vertices k-means algorithm. empirically demonstrate subtree kernel, is prominent kernels, our significantly outperforms state-of-the-art in terms datasets contain structurally complex graphs beyond typically considered molecular graphs.