Comparison between various regression depth methods and the support vector machine to approximate the minimum number of misclassifications (Q1855638): Difference between revisions

This paper deals with the following problem: from a given set of observations \(z_{n}=\{(x_{i,1},\dots,\) \(x_{i,p-1},y_{i})\); \(i=1,\dots,n\}\subset R^{p}\), where \(x_{i}=(x_{i,1},\dots,x_{i,p-1})\in R^{p-1}\), \(y_{i}\in\{0,1\},\;i=1,\ldots,n\), are responses, find an affine hyperplane defined via \(\theta\in R^{p}\) such that a good classification of the responses is possible. A central role concerning the existence and the quality of such estimates plays the quantity \(n_{co}\) defined as the minimum number of misclassifications, for the given data set \(z_{n}\), that any affine hyperplane must incur. The authors introduce two new methods for finding reasonable approximation procedures for the minimum number of misclassifications. The proposed methods are modifications of the regression depth method. They employ the support vector machine in a way especially adapted to the regression depth problem. None of the considered approximation algorithms outperform all others, but the authors give some recommendations how to choose a reasonable algorithm. Pseudo code for a heuristic method is presented.