Dynamical Feature Relevance Estimation Using Goldman Typical Testors

The estimation of feature relevance in the framework of supervised classification problems has a great practical significance. Nevertheless, to solve this problem in real situations it is not always an easy task. One of the complex situations appears when the result of comparison between two coordinate values of the description of the respective objects is a real number. Goldman typical testors are useful for estimating feature relevance in supervised classification problems where feature values are compared using real criteria. Truly the computational complexity of all of these testor algorithms is too high. In real world problems very frequently occur modifications of the training matrix. Any modification to the training matrix can change the set of all Goldman typical testors, and this set must be computed again after each modification. In this paper we analyze how does the set of all Goldman typical testors of a comparison matrix change after modifications. An alternative method to calculate all Goldman typical testors of the modified matrix, more efficient than any traditional algorithm, is exposed. The new method’s complexity is analyzed and some experimental results are showed.