Detecting Similar Programs via The Weisfeiler-Leman Graph Kernel

With the increasing availability of source code on the Internet, many new approaches to retrieve, repair, and reuse code have emerged that rely on the ability to efficiently compute the similarity of two pieces of code. The meaning of similarity, however, heavily depends on the application domain. For predicting API calls, for example, programs can be considered similar if they call a specific set of functions in a similar way, while for automated bug fixing, it is important that similar programs share a similar data-flow. In this paper, we propose an algorithm to compute program similarity based on the Weisfeiler-Leman graph kernel. Our algorithm is able to operate on different graph-based representations of programs and thus can be applied in different domains. We show the usefulness of our approach in two experiments using dataflow similarity and API-call similarity.