Distance-Based Dynamic Weight: A Novel Framework for Multi-Source Information Fusion
Cuiping Cheng, Xiaoning Zhang, Taihao Li · 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) · 2022
Belief theory have been widely used in multi-source information (MSI) fusion due to its advantages in modeling uncertainty. However, combining highly conflicting MSI still remains a big challenge for belief theory. In this paper, we present a novel dynamic weight-based fusion framework, called DwFuse, which can effectively deal with conflicting data by dynamically integrating different evidences at the decision level. Specifically, we first formulate a dynamic weight generation function to explore the optimal contribution which uses evidence distance as a conflict measurement. Unlike most previous work only focusing on linear association of weight and distance, we develop a nonlinear function by joint distance and weight variables, which can better characterize the dynamic properties of complex systems. Then, by integrating the obtained dynamic weights into the Dempster-Shafer combination rule, we propose a general multimodel-based belief generation architecture and extensive experiments have been done on large corpus. The results demonstrated that our framework is more robust and effective compared with others on real datasets.