Group for Research in Decision Analysis

G-2017-33

No-means clustering: A Stochastic variant of \(k\)-means

, , and

Simple, intuitive, and scalable to large problems, \(k\)-means clustering is perhaps the most frequently-used technique for unsupervised learning. However, global optimization of the \(k\)-means objective function is challenging, as the clustering algorithm is highly sensitive to its initial value. Exploiting the connection between \(k\)-means and Bayesian clustering, we explore the benefits of stochastic optimization to address this issue. Our "\(\texttt{no means}\)" algorithm has provably superior mixing time to a natural Gibbs sampler with auxiliary cluster centroids. Yet, it retains the same computational complexity as the original \(k\)-means approach. Comparisons on two benchmark datasets indicate that stochastic search usually produces more homogeneous clusters than the steepest descent algorithm employed by \(k\)-means. Our \(\texttt{no means}\) method objective function has multiple modes which are not too far apart.

, 13 pages