Multilevel spectral hypergraph partitioning with arbitrary vertex sizes

This paper presents a new spectral partitioning formulation which directly incorporates vertex size information by modifying the Laplacian of the graph. Modifying the Laplacian produces a generalized eigenvalue problem, which is reduced to the standard eigenvalue problem. Experiments show that the scaled ratio-cut costs of results on benchmarks with arbitrary vertex size improve by 22% when the eigenvectors of the Laplacian in the spectral partitioner KP are replaced by the eigenvectors of our modified Laplacian. The inability to handle vertex sizes in the spectral partitioning formulation has been a limitation in applying spectral partitioning in a multilevel setting. We investigate whether our new formulation effectively removes this limitation by combining it with a simple multilevel bottom-up clustering algorithm and an iterative improvement algorithm for partition refinement. Experiments show that in a multilevel setting where the spectral partitioner KP provides the initial partitions of the most contracted graph, using the modified Laplacian in place of the standard Laplacian is more efficient and more effective in the partitioning of graphs with arbitrary-size and unit-size vertices; average improvements of 17% and 18% are observed for graphs with arbitrary-size and unit-size vertices, respectively. Comparisons with other ratio-cut based partitioners on hypergraphs with unit-size as well as arbitrary-size vertices, show that the multilevel spectral partitioner produces either better results or almost identical results more efficiently.