An approach of classifiers fusion based on hierarchical modifications

Classifiers fusion is considered as an effective way to promote the accuracy of pattern recognition. In practice, its performance is mainly limited by potentials and reliabilities of base classifiers, which are learned from different attribute spaces. In order to overcome the above problems, we present a new approach of classifiers fusion based on hierarchical modifications in the framework of belief function theory. At first, an intra-attribute modification is proposed to taking into account the potentials and reliabilities of base classifiers. Instead of discounting a classifier with a weight only, we employ a piece of evidence derived from the nearest labeled neighbor to modify the weighted output of one base classifier in its individual attribute space. Then, the modified output is combined with other modified results from their own attribute spaces and this procedure could be seen as an inter-attribute modification. Both modifications aim to make the classification result as close to the truth as possible, so we take them into account to construct a new objective function for optimizing the weights. Finally, some real data sets are used in experimental applications to demonstrate that the proposed method is superior to other related belief based fusion methods.