A new probability adjustment method for combining conflicting evidences: application in classifiers combination

Abstract Dempster-Shafer (DS) evidence theory has been widely used in information fusion. However, combining conflicting evidences is still an open question despite the existence of many proposed studies to improve the DS theory combination rule. In this paper we propose an improved probability adjustment method to combine evidences based on their ranks. The idea of our method is to divide each piece of evidence by its rank, assuming that the evidences produced by each source of information are sorted descendingly. The obtained evidences are normalized, averaged and then the DS combination rule is applied. The main objective of our method is to employ the evidence rank to adjust it by giving high support to highly ranked evidences and low support to lowly ranked evidences. Through a numerical example, we proved that the proposed method is more efficient and generates better results that those achieved by many other DS based combination methods. Moreover, we applied our method in combining the outputs generated by multiple single-label classifiers. Using many real-world classification datasets, we compared our method with other DS combination methods, non-learning combination rules and ensemble classifiers. The achieved results shows that our proposed method performs better using most of the datasets.