Title: Combining diverse classifiers using precision index functions

Authors: Jose Bird; Daijin Ko

Addresses: Department of Management and Statistics, The University of Texas at San Antonio, One UTSA Circle, San Antonio, Texas 78249, USA ' Department of Management and Statistics, The University of Texas at San Antonio, One UTSA Circle, San Antonio, Texas 78249, USA

Abstract: This paper introduces a combining classifier using two proposed precision indexes: precision index (PIN) and class specific precision index (PIC). Comparison of combining methods typically fails to consider high precision performance. This new combining method generates predictions with higher precision and recall than other methods. The proposed method is especially useful for efficient screening of predictions where actual verification is time consuming and costly. The performance of the proposed method is compared to majority voting, stacking, and cluster-selection for two well-known datasets: 1) vowel recognition (Hastie et al., 2009); 2) yeast protein localisation (Frank and Asuncion, 2010). The precisions obtained exceeded results previously reported for protein localisation data (Horton and Nakai, 1997; Chen, 2010) and for vowel recognition data (Hastie et al., 2009). A weighted precision index using PIC and PIN indexes outperformed all combining methods at higher precisions.

Keywords: precision index; PIN; class-specific precision index; PIC; combined classifiers; recall; pattern recognition; prediction screening; vowel recognition; yeast protein localisation; precision performance.

DOI: 10.1504/IJAPR.2013.052338

International Journal of Applied Pattern Recognition, 2013 Vol.1 No.1, pp.3 - 26

Received: 19 Aug 2012
Accepted: 14 Sep 2012
Published online: 31 Jul 2014 *

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