A theoretical and experimental analysis of linear combiners for multiple classifier systems

• DOI: 10.1109/tpami.2005.109
• Publication Date: 2005-04-25T18:49:32Z
• Authors: G. Fumera; F. Roli
• Keywords: Linear classifier

In this paper, a theoretical and experimental analysis of linear combiners for multiple classifier systems is presented. Although are the most frequently used combining rules, many important issues related to their operation pattern classification tasks lack basis. After critical review framework developed in works by Tumer Ghosh on which our based, we focus simplest widely implementation combiners, consists assigning nonnegative weight each individual classifier. Moreover, consider ideal performance rule, i.e., that achievable when optimal values weights used. We do not problem estimation, has been addressed literature. Our shows how terms misclassification probability, depends classifiers, correlation between outputs. particular, evaluate improvement can be achieved using weighted average over simple rule investigate what way it classifiers. Experimental results real data sets show behavior agrees with predictions analytical model. Finally, discuss contribution state art practical relevance systems.