A jensen-shannon kernel for hypergraphs

Abstract

In this paper we explore how to construct a Jensen-Shannon kernel for hypergraphs. We commence by calculating probability distribution over the steady state random walk on a hypergraph. The Shannon entropies required to construct the Jensen-Shannon divergence for pairs of hypergraphs are obtained from steady state probability distributions of the random walk. The Jensen-Shannon divergence between a pair of hypergraphs is the difference between the Shannon entropies of the separate hypergraphs and a composite structure. Our proposed kernel is not restricted to hypergraphs. Experiments on (hyper)graph datasets extracted from bioinformatics and computer vision datasets demonstrate the effectiveness and efficiency of the Jensen-Shannon hypergraph kernel for classification and clustering. © 2012 Springer-Verlag Berlin Heidelberg.