A fuzzy preference-based Dempster-Shafer evidence theory for decision fusion

Dempster-Shafer evidence theory (D-S) is an effective instrument for merging the collected pieces of basic probability assignment (BPA), and it exhibits superiority in achieving robustness of soft computing and decision making in an uncertain and imprecise environment. However, the determination of BPA is still uncertain, and merely applying evidence theory can sometimes lead to counterintuitive results when lines of evidence conflict. In this paper, a novel BPA generation method for binary problems called as the base algorithm is designed based on the kernel density estimation to construct the probability density function models, using the pairwise learning method to establish binary classification pairs. By means of the new BPA generation method, a new decision-making algorithm based on D-S evidence theory, fuzzy preference relation and nondominance criterion is effectively designed. The strength of the proposed method is presented in applying pairwise learning, which transforms the original complex problem into simple subproblems. With this process, the complexity of the problem to be solved is greatly reduced, which increases the feasibility for industrial applications. Furthermore, the fuzzy computing technique is used to aggregate the output for each single subproblem, and the nondominance degree of each class is determined from the fuzzy preference relation matrix, which can be directly used for the determination of the input instance. Based on several industrial-based classification experiments, the proposed BPA generation method and decision-making algorithm present the effectiveness and improvement in terms of precision and Cohen's kappa