Constructing basic belief assignment from feature data

The Dempster-Shafer Theory of Evidence provides a set of effective tools to model the belief induced by uncertain evidence, which is widely used in information fusion area. The Basic Belief Assignment (BBA) is a fundamental function within the theory. However, it does not offer any universal mechanism to construct BBA. The construction of BBA is also referred as the modeling of evidence. A novel feature-based method for modeling of evidence is proposed in this paper. The focal elements consist of each class label in the frame of discernment and the frame of discernment itself. A referenced centric vector and several scattered parameters are set to each of them. Those centric vectors and parameters are obtained optimally through supervised learning. If a new feature is input, its Euclidean distances to each referenced vector are utilized to derive a BBA. The proposed method can be efficiently implemented with few learning samples. To account for a classification task, it can differentiate either uncertainty or ignorance that leads to rejection. Several experiments are performed to demonstrate its efficiency and rationality. The simulation of a fusion target recognition scenario also indicates it is a promising tool compared with classical Bayesian method.