The Weisfeiler-Lehman Kernel

15 years. One reason for their attractiveness lies in the fact that the underlying theory of these algorithms can
easily be generalised from vectorial data to strings, time series, graphs and other types of
structured data. In real-world applications, the efficient computation of the kernel
function, i.e. of the similarity measure, is a key challenge. For strings and time series, efficient computation
techniques for kernels were developed early on, but graph kernels remained slow to compute and only applicable to
graphs with a few dozen nodes without attributes. A major focus of our research has been to turn graph kernels from a theoretical concept into a useful
tool for practical graph data analysis. In this talk, we will present our work on efficient graph kernels, in
particular a recent breakthrough from 2009, which now allows for highly scalable graph kernel computation (Shervashidze, Borgwardt. Fast
Subtree Kernels on Graphs, NIPS Outstanding Student Paper Award 2009).