Data-and knowledge-driven belief rule learning for hybrid classification

Abstract

In hybrid classification problems, apart from labeled data, some related expert knowledge may also be obtained. If the partial information from these two sources can be jointly used well, the performance may be effectively improved. To this end, we develop a new framework for hybrid classification by modeling the uncertain data and knowledge in the form of belief rules under the framework of belief function theory. First, a data-driven belief rule learning algorithm is proposed to learn a compact and interpretable belief rule model from labeled training data. Then, a knowledge-driven belief rule learning algorithm is proposed to learn a model which is complementary to that learned from data via active learning strategy. Finally, the genetic integration of belief rules is developed in order to integrate data-driven belief rules with knowledge-driven ones to reduce the rule redundancy and rule conflict by considering both classification accuracy and model interpretability. Experiments based on both simulated and real data sets demonstrate the superiority of the proposed model for integrating data and knowledge.