A Novel Information Fusion Approach Based On Group Decision-making And Evidence Theory And Its Application To Multiple Classifiers Ensemble

As an extension to the classical probability theory, the evidence theory based on Bayesian theory has been developed to be an uncertainty reasoning method which reflects the degree of belief on the power set of all possible results with upper probabilities and lower probabilities, and gives the results more conformed to people's habits of thinking. Also the evidence theory can reasonably describe and process nonconflicting "incomplete", "uncertain" and "unsure" information, and has been successfully used to solve a variety of uncertainty information problems from many areas. However, it often encounters a difficulty of uncertain conflicting information in real applications. Especially, when bigger differences or conflicts exist, its combination rule is even not useable or gives a contradicting result. To a large extent, this situation restricts the further development and applications of evidence theory, which should be further studied and necessarily extended to greater applications.In this paper, we attempt to introduce the idea of group decision-making into the theory of evidence, propose a new combination rule for conflicting evidences, and apply to information fusion of multiple classifiers. The main research works are as follows:1. Consider a recent study on the conflict evidence combination, the robust combination rule as a generalization to classical combination rules. It can not only avoid conflicting problems but also has some good properties such as quasi-associative and robustness. However, it has not the consistency of focal elements and satisfactory combination results or missing much information sometimes. To aim at its disadvantages, we propose a new rule for conflicting evidences, adaptive-robust combination rule. The effectiveness of the proposed rule is illustrated and compared by examples.2. Each combination rule of classical evidence theory uses only one kind of decision-making systems as default. But indeed different situations correspond to different decision-making systems. In this situation, we apply the main rules of group decision-making to preprocess conflicting evidences to be combined. A novel information fusion method is presented according to the improved rule. Experiments show that the results obtained by the method are more rational and applicable.3. To increase classification accuracy of classifical multiple classifiers, we integrate our information fusion method with Fisher discriminant analysis. Different classifiers generated according to-different attributes are fused so as to give the integrated classification result with higher accuracy. Usability and superiority of the new method are demonstrated further.