A New Probability Adjustment Method for Combining Conflicting Evidences: Application in Classifier Combination

We propose in this paper an improved probability adjustment method to combine multiple single-label classifiers using the Dempster–Shafer (DS) combination rule. Our method employs all types of classifier outputs namely the class label, its score (also called evidence or support score), and its rank. The idea of our method is to adjust the score value assigned by a given classifier using its rank. Our method aims to assign high support to highly ranked labels and low support to lowly ranked labels. In contrast with other DS-based combination methods, our method does not require a complicated process to adjust the pieces of evidence. Through a numerical example, we proved that the proposed method is efficient and generates better results than those achieved by other DS-based combination methods. Moreover, we applied our method in combining multiple classifiers using many real-world datasets. The comparison of our method with other DS-based combination methods, non-learning combination methods, and ensemble classifiers shows that our method performs better using most of the datasets.