Asynchronous optimization approach for evidential reasoning rule-based classifier

Abstract

Data classification is an important and challenging issue encountered in many practical applications. The classifier based on Evidential Reasoning Rule (ER Rule) can well handle the uncertainty in classification and obtain competitive accuracy. However, there are many parameters to be optimized, and the computation cost of the usually adopted optimization strategy is relatively high. This fact weakens the application advantage of ER Rule classifier. To address this challenge, an asynchronous optimization approach is proposed. In the original ER Rule classifier, feature referential values and evidence weights are optimized synchronously; while in the proposed method, these two types of parameters are optimized separately based on their essential impacts on the classification results. For the optimization of feature referential values, the objective function measures the quality of belief matrix, which is the critical strategy to improve the computational efficiency. Three types of feasible objective functions are constructed. After obtaining optimal referential values, evidence weights are optimized based on the actual classification effect, and the required iterations decrease. Various experiments on 14 publicly available datasets verify the computational efficiency and classification performance of the proposed method.