The Use of Collective Method for Improvement of Regression Modeling Stability

by Oleg V. Senko.

Abstract: The collective methods of regression analysis are discussed in the paper as the tools to improve the mean discrepancy of regression modeling. The mean discrepancy in the paper is calculated not only by the space of multidimensional observations but also by the space of training sets of fixed size. It was shown that mean discrepancy may be represented as the sum of three component. The first one is irremovable "noise", the second component is the mean squared deviation of mean regression function from conditional means of dependent variable in points of independent variables space, the third component is instability term that is the variance of regression functions on the space of training sets. It was shown that the using of simplest voting procedure by the initial set of regression methods allows to receive new regression model with instability component less than average instability component of the initial set .The large scale experiments with Monte-Carlo simulated data demonstrated that voting procedures really allow to improve regression performance and that the cause of such improvement is the decrease of instability of training.

Oleg V. Senko,

Editor: Chaganty, N. Rao,

READING THE ARTICLE: You can read the article in portable document (.pdf) format (188141 bytes.)

NOTE: The content of this article is the intellectual property of the authors, who retains all rights to future publication.

This page has been accessed 2697 times since July 24, 2006.

Return to the InterStat Home Page.