Title: A graphic method for choosing the best parameter value of margins for maximising in GA-Ensemble

Authors: Dong-Yop Oh; J. Brian Gray

Addresses: Computer Information Systems and Quantitative Methods Department, The University of Texas-Pan American, 1201 W University Drive, Edinburg, TX 78539-2999, USA ' Department of Information Systems, Statistics and Management Science, The University of Alabama, Tuscaloosa, AL 35487-0226, USA

Abstract: We examine how to choose the pth percentile of margins of a fitness function for GA-Ensemble and provide a new approach for an outlier diagnostic method based on an appropriately chosen pth percentile of margins in GA-Ensemble. This provides a solution from the optimised set of base classifiers with their weights using a genetic algorithm for a binary classification problem. A plot of margins vs. maximised percentile to select p is employed to decide which p to use through the gap among the margins of outliers and normal examples based on a synthetic data set as well as real data sets. GA-Ensemble using the best pth percentile of margin is superior to AdaBoost and GA-Ensemble using other pth percentile of margin for two real-world data sets from the UC-Irvine Machine Learning Repository.

Keywords: binary classification; GA ensemble; genetic algorithms; margins; outliers; fitness function.

DOI: 10.1504/IJSS.2015.070689

International Journal of Services and Standards, 2015 Vol.10 No.3, pp.89 - 102

Received: 27 Feb 2014
Accepted: 01 Dec 2014
Published online: 19 Jul 2015 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article