Big-data feature screening using bregman divergence

Modern biomedical data usually are big: they have high dimensions and/or massive volumes as a result of fast advancement in sensing and information technologies. Many components of the data, however, may not play any important role for tasks of interest. Reducing the dimensionality of the data turns out to be critical for obtaining accurate result and a good generalization capability. This paper introduces an approach for using Bregman divergence to describe the discriminating capability of a subset of features. An important advantage of this method is that the nonlinear relationship existing in the data can be exploited with the Bregman divergence. Moreover, the optimization of the proposed discrimination measure turns out to be particularly simple, which leads to a fast algorithm.