Classifier fusion based on cautious discounting of beliefs

Classifier fusion is a classical approach to improve the classification accuracy. The multiple classifiers to combine have in general different classification qualities (i.e. performances), and the proper evaluation of the classifier quality plays an important role for achieving the best global performance. We propose a new method for classifier fusion based on refined reliability evaluation (CF-RRE). For each object, the reliability of its classification result with a given classifier is characterized by a matrix Rc×c (c being the number of classes in the data set), which is estimated based on the classifier performance in the neighborhoods (i.e. K-nearest neighbors) of the object using the training data. The reliability matrix is used to make a cautious discounting of the classification result. More specifically, the probability (or belief) of the object associated with each class is cautiously redistributed according to the reliability matrix under the belief functions framework. The discounted classification results of each classifier can be combined by Dempster's rule for making the final class decision. Our simulation results illustrate the potential of this new method using real data sets, and they show that CF-RRE can improve substantially the classification accuracy.