Selection and statistical validation of features and prototypes

Abstract

Features and protypes selection are two major problems in data mining, especially for machine learning algorithms. The goal of both selections is to reduce storage complexity, and thus computational costs, without sacrificing accuracy. In this article, we present two incremental algorithms using geometrical neighborhood graphs and a new statistical test to select, step by step, relevant features and prototypes for supervised learning problems. The feature selection procedure we present could be applied before any machine learning algorithm is used.