Path optimization and near-greedy analysis for graph partitioning: an empirical study
This paper presents the results of an experimental study of graph partitioning. We describe a new heuristic technique, path optimization, and its application to two variations of graph partitioning: the max_cut problem and the min_quotient_cut problem. We present the results of computational comparisons between this technique and the Kernighan-Lin algorithm, the simulated annealing algorithm, the FLOW-lagorithm the multilevel algorithm, and teh recent 0.878-approximation algorithm. The experiments were conducted on two classes of graphs that have become standard for such tests: random and random geometric. They show that for both classes of inputs and both variations of the problem, the new heuristic is competitive with the other algorithms and holds an advantage for min_quotient_cut when applied to very large, sparse geometric graphs. In the last part of the paper, we describe an approach to analyzing graph partitioning algorithms from the statistical point of view. Every partitioning of a graph is viewed as a result achieved by a "near gready" partitioning algorithm. The experiments show that for "good" partitionings, the number of non-greedy steps needed to obtain them is quite small; moreover, it is "statistically" smaller for better partitionings. This led us to conjecture that there exists an "optimal" distribution of the non-greedy steps that characterize the classes of graphs that we studied.