The Sum-over-Forests clustering

This work introduces a novel way to identify dense regions in a graph based on a mode-seeking clustering technique, relying on the Sum-Over-Forests (SoF) density index [1] (which can easily be computed in closed form through a simple matrix inversion) as a local density estimator. We rst identify the modes of the SoF density in the graph. Then, the nodes of the graph are assigned to the cluster corresponding to the nearest mode, according to a new kernel, also based on the SoF framework. Experiments on arti cial and real datasets show that the proposed index performs well in nodes clustering.