A Mixed Weisfeiler-Lehman Graph Kernel

Abstract

Using concepts from the Weisfeiler-Lehman (WL) test of isomorphism, we propose a mixed WL graph kernel (MWLGK) framework based on a family of efficient WL graph kernels for constructing mixed graph kernel. This family of kernels can be defined based on the WL sequence of graphs. We apply the MWLGK framework on WL graph sequence taking into account the structural information which was overlooked. Our MWLGK is competitive with or outperforms the corresponding single WL graph kernel on several classification benchmark data sets.