An adaptive approach to non-parametric estimation of dynamic probability density functions

Abstract

Accurate and flexible probability density estimation is fundamental in machine learning tasks, in classification and routine data analyses applications. In this paper we propose an adaptive version of the Histogram Trend Filtering (HTF), which is a relatively new method used for non-parametric density estimation. This technique enjoys low computational complexity, while being able to automatically detect abrupt changes in the underlying dynamics of the estimated distribution. Therefore, it can deal with estimating both stationary and non-stationary distributions.