Defining Roles in Transaction Networks Using Deep Learning

Abstract

Recent years have seen an immense increase in the amount of data our generated by humans. This data takes many different shapes and forms, and one of the types we find most often is in the form of network data. In order to be able to extract valuable insights from these data, we need robust tools to analyze these networks. An area of research that has received a lot of attention recently, has been finding methods that can be used to extract roles from graph data.

In this thesis, we examine an algorithm called struc2vec that embeds a graph into a vector space and how we can use these embeddings to extract roles from the graph. We consider embeddings of various dimensionalities and how the choice of dimensionality impacts how well the embeddings represent the roles in the graph. Where applicable, we use dimensionality reduction techniques to obtain lower-dimensionality embeddings from which we can extract the roles using clustering methods.

Our results show that struc2vec generates embeddings that separate nodes well with regard to their role, and manages to capture several aspects of the graph structure.