Measuring the expressivity of graph kernels through the rademacher complexity

Graph kernels are widely adopted in real-world applications that involve learning on graph data. Different graph kernels have been proposed in literature, but no theoretical comparison among them is present. In this paper we provide a formal definition for the expressiveness of a graph kernel by means of the Rademacher Complexity, and analyze the differences among some state-of-the-art graph kernels. Results on real world datasets confirm some known properties of graph kernels, showing that the Rademacher Complexity is indeed a suitable measure for this analysis.