Machine learning algorithm based on convex hull analysis

Abstract

In this paper machine learning methods for automatic classification problems using computational geometry are considered. Classes are defined with convex hulls of points sets in a multidimensional feature space. Classification algorithms based on the estimation of the proximity of the test point to convex class shells are considered. Several ways of such estimation are suggested when the test point is located both outside the convex hull and inside it. A new method for estimating proximity based on linear programming is proposed, and the corresponding nearest convex hull classifier is described. The results of experimental studies on the real medical diagnostics problem are presented. An efficiency comparison of the proposed classifier and other types of classifiers, both based on convex hull analysis and not, has shown the high efficiency of the proposed method for estimating proximity based on linear programming.