Speed Optimizations to Parzen Window Classifier Using Probability Approximation, Discretization and Compression

Abstract

Parzen window estimators can model any type of complicated probability density manifolds. However, when it comes to real life applications, they are not as popular as the Artificial Neural Networks or the Support Vector Machines. That is mainly because Parzen window classifiers require long and complex calculations during the classification process. This article introduces speed optimization methods for Parzen window classifier that makes this classifier faster than any other convergent classifier at a small performance cost. The method includes, discretization, look-up tables, approximation, and probabilistic compression. Experiments conducted on both computer generated and real-life data prove that the resultant classifier is only slightly less accurate than Artificial Neural Networks and Support Vector Machines while immensely faster.