Label Propagation for Hypergraph Partitioning

Many problems in computer science can be represented by a graph and reducedto a graph clustering or k-way partitioning problem. In the classical definition,a graph consists of nodes and edges which usually connect exactly two nodes.Hypergraphs are a generalization of graphs, where every edge can connect anarbitrary number of nodes. Recent results suggest that some problems in com-puter science are better and more intuitively modeled with hypergraphs insteadof graphs. This thesis investigates the adaptation of label propagation, a graphclustering algorithm, to hypergraph partitioning. We propose three adaptationsof label propagation which are motivated by graph-based hypergraph modelingand evaluate them as coarsening strategies in a direct k-way multilevel hyper-graph partitioning framework. Furthermore, we propose a greedy local searchalgorithm inspired by label propagation for the uncoarsening and refinementphase of the multilevel partitioning heuristic. We compare our algorithms to thestate-of-the-art hypergraph partitioners hMetis and PaToH. Our results implythat the utilization of label propagation in the multilevel hypergraph partition-ing scheme is promising, as we outperform both hMetis and PaToH on VLSIinstances for larger values of k: for k = 128 our proposed algorithms produce 2%better cuts than hMetis and 4% better cuts than PaToH.