K-Medoids For K-Means Seeding

Abstract

We run experiments showing that algorithm clarans (Ng et al., 2005) finds better K-medoids solutions than the Voronoi iteration algorithm. This finding, along with the similarity between the Voronoi iteration algorithm and Lloyd's K-means algorithm, suggests that clarans may be an effective K-means initializer. We show that this is the case, with clarans outperforming other seeding algorithms on 23/23 datasets with a mean decrease over K-means++ of 30% for initialization mse and 3% or final mse. We describe how the complexity and runtime of clarans can be improved, making it a viable initialization scheme for large datasets.

v1: (24 pages, 9 figures) v2: not at 33-rd ICML: forgot to modify .sty file. Reordered sections. Simplified to be specific to K-means seeding. New experiments. v3: (22 pages, 10 figures) Modified .sty file. Minor cosmetic changes. v4: added references and disussion of 2 related works v5: NIPS camera ready