Density boosting for Gaussian mixtures

Ensemble method is one of the most important recent developments in supervised learning domain. Performance advantage has been demonstrated on problems from a wide variety of applications. By contrast, efforts to apply ensemble method to unsupervised domain have been relatively limited. This paper addresses the problem of applying ensemble method to unsupervised learning, specifically, the task of density estimation. We extend the work by Rosset and Segal [3] and apply the boosting method, which has its root as a gradient descent algorithm, to the estimation of densities modeled by Gaussian mixtures. The algorithm is tested on both artificial and real world datasets, and is found to be superior to non-ensemble approaches. The method is also shown to outperform the alternative bagging algorithm.