Likelihood-based Clustering of Directed Graphs

In this paper, a new, stochastic approach to the clustering of directed graphs is presented. This method differs from the commonly used ones by defining the term "cluster" in an alternative way: a cluster can even be a set of vertices that don't connect to each other at all, provided that they have the same connectional preference to other vertices. First, a short overview of the current state of the art will be given. Then the underlying theory of this alternative clustering method will be explained and a possible implementation will be proposed. To support the validity of this approach, benchmark results on computer-generated graphs as well as two real applications are presented.