A0357
Title: Power k-means clustering
Authors: Jason Xu - Duke University (United States) [presenting]
Abstract: An alternative to Lloyd's classic algorithm for k-means clustering is proposed that retains its simplicity but mitigates its tendency to get trapped by local minima. Called power k-means, the method embeds the k-means problem in a continuum of similar, better behaved problems with fewer local minima. The previously discovered k-harmonic means algorithm coincides with one point along this continuum. Power k-means anneals its way toward the solution of ordinary k-means by way of majorization-minimization (MM), sharing the appealing descent property and low complexity of Lloyd's algorithm. Further, the method complements widely used seeding strategies, reaping marked improvements when used together. We demonstrate its advantages over simulated and real data examples.
