Evaluating the reliability of sources of evidence with a two-perspective approach in classification problems based on evidence theory

Abstract

Conflict management and accuracy improvement are the two main concerns of classification problems based on the evidence theory. High conflict among sources of evidence can be solved effectively using discounting methods based on source-reliability evaluations. However, these methods may not ensure efficient performance of a classification model. To relieve high conflict and improve accuracy, a two-perspectives approach for reliability evaluation is presented to generate discounting rules. An independent reliability evaluation (IRE) is used to assess the independent reliability of an individual source, under the assumption that the source works independently. The other perspective is the combination reliability evaluation (CRE). It evaluates all the sources by considering the combination relationship among them. Both methods are designed as supervising methods and integrate a new dissimilarity measure proposed in this paper—decision dissimilarity—with the Jousselme distance. The ability of the new dissimilarity measure to effectively discriminate evidence from the truth can be experimentally verified. The proposed approach is not only effective for conflict management but also for the improvement of the performance of classification models based on the evidence theory as it helps implement the correct and specific decisions.