BNAK-Divide-and-Merge Clustering Algorithm

Divide-and-Merge is a methodology for clustering a set of objects that combines a top-down "divide" method with a bottom-up "merge" method. In this paper, we propose a 2-way normalized cut with automatically determining K clustering algorithm (BNAK-Divide-and-Merge) based on the Divide-and-Merge. In order to improve the efficiency and performance of the divide phase, our methodology alternately uses 2-way normalized cut spectral clustering algorithm with a threshold to limit the number of tree nodes produced by the divide phase. Furthermore, we present a measurement of automatically determining the expected number of clusters (i.e., K) at the merge phase so that it not only reduces the number of additional parameters which must be inputted manually, but also allows the algorithm to control the clustering quality. We also give empirical results on four common well-known data sets where the algorithm performs better than or competitively with k-means and Divide-and-Merge.