Neighborhood Approximate Reducts-Based Ensemble Learning Algorithm and Its Application in Software Defect Prediction

Ensemble learning is a machine learning paradigm that integrates the results of multiple base learners according to a certain rule to obtain a better classification result. Ensemble learning has been widely used in many fields, but the existing methods still have the problems of difficult to guarantee the diversity of base learners and low prediction accuracy. In order to overcome the above problems, we considered ensemble learning from the perspective of attribute space division, defined the concept of neighborhood approximate reduction through neighborhood rough set theory, and further proposed an ensemble learning algorithm based on neighborhood approximate reduction, called ELNAR. ELNAR algorithm divides the attribute space of the data set into multiple subspaces. The basic learners trained based on the data sets corresponding to different subspaces have great differences, so as to ensure the strong generalization performance of the ensemble learner. In order to verify the effectiveness of ELNAR algorithm, we applied ELNAR algorithm to software defect prediction. Experiments on 20 NASA MDP data sets show that ELNAR algorithm can better improve the performance of software defect prediction compared with the existing ensemble learning algorithms.