Software defect number prediction (SDNP) is a primary field of study in software engineering, with the goal of optimizing the test resources allocation by predicting the number of defects in software modules. In recent years, ensemble learning has been widely employed in SDNP. However, how to enhance the diversity of base learners remains a challenge in ensemble learning. In this paper, we first introduce the notion of neighborhood granularity entropy (abbreviated to NGE), which combines neighborhood entropy and neighborhood knowledge granularity within the framework of neighborhood rough sets. Second, we propose a novel feature selection algorithm called neighborhood granularity entropy-based feature selection (abbreviated to NGEFS) based on NGE. Third, we combine the NGEFS algorithm with the traditional random subspace selection (abbreviated to RSS) technique and propose the notion of neighborhood random super-reduct (abbreviated to NRSR). Moreover, we propose an ensemble learning algorithm, called ELNRSR, based on NRSR. Finally, we use ELNRSR to predict the number of defects in software modules. To address the class-imbalanced problem, we combine ELNRSR with the SMOTER (synthetic minority oversampling technique for regression) algorithm. Experimental results on 22 datasets demonstrate that ELNRSR significantly enhances the performance of SDNP compared to existing ensemble learning methods.