Dynamic Combination of Multiple Classifiers Based on Normalizing Decision Space
Jiangtao Huang, Minghui Wang · 2010
This paper presents a novel fusion method for final decision making in multiple classifier system. The novel approach proposed fuses multiple individual classifiers through normalizing and re-subdividing classification output space. Firstly, each class decision output of each classifier in multiple classifier system is divided into four spaces by reliability of corresponding classifier. And then, reject degree is used to indicate classification to substitute for original output. Secondly, dynamic probability of each reject degree which used to combine multiple classifiers is calculated by using synchronization decision support value defined in this paper. Finally, the final decision can be made by integrated support degree based on support space of all classes. Experimental results demonstrate that the method is able to achieve a preferable solution, which has a better classification performance compared to single classifier and other combination methods.