A hierarchical clustering algorithm based on the Hungarian method

We propose a novel hierarchical clustering algorithm for data-sets in which only pairwise distances between the points are provided. The classical Hungarian method is an efficient algorithm for solving the problem of minimal-weight cycle cover. We utilize the Hungarian method as the basic building block of our clustering algorithm. The disjoint cycles, produced by the Hungarian method, are viewed as a partition of the data-set. The clustering algorithm is formed by hierarchical merging. The proposed algorithm can handle data that is arranged in non-convex sets. The number of the clusters is automatically found as part of the clustering process. We report an improved performance of our algorithm in a variety of examples and compare it to the spectral clustering algorithm.