Improving clustering algorithms through constrained convex optimization

Inspired by the recent successes of boosting algorithms, a trend in unsupervised learning has begun to emphasize the need to explore the design of weighted clustering algorithms. We handle clustering as a constrained minimization of a Bregman divergence. Theoretical results show benefits resembling those of boosting algorithms, and bring new modified weighted versions of clustering algorithms such as k-means, expectation-maximization (EM) and k-harmonic means. Experiments display the quality of the results obtained, and corroborate the advantages that subtle data reweightings may indeed bring to clustering.