Extended belief rule-based system with online joint learning strategy

Abstract

With the proliferation of dynamic data streams in the advanced technology environment, it is necessary to solve the evolving classification problems by using adaptable and interpretable artificial intelligence techniques. To meet this challenge, a new extended belief rule-based (EBRB) system incorporating online joint learning strategy is proposed in this paper. The online joint learning strategy comprises two key components: rule update and parameter update schemes. In the rule update scheme, different rule incorporation processes are designed for the labeled or unlabeled input data while overlapping and redundant rules are removed from the rule base. To adapt the updated rule base, the parameter update scheme is designed to retune the parameters within updated rule base. The antecedent attribute weights are optimized using the Bayesian optimization algorithm and the rule weights are updated based on the consistency of rules. To evaluate the performance of the developed system, it is applied to assist radiologists in diagnosing thyroid nodules. Compared with the existing offline EBRB systems and online learning methods, the proposed online joint learning EBRB system could generate higher classification accuracy with fewer rules in the limited running time.