Software defect prediction (SDP) is a primary field of study in software engineering, aiming to optimize test resource allocation by highlighting the defect-prone software modules. Over the last few years, ensemble learning method has been extensively adopted in SDP. However, how to strengthen the diversity of base learners is an issue in ensemble learning. In this paper, we consider the problem of diversity in the eye of feature space perturbation. First, we propose the notion of neighborhood granularity discrimination index (NGDI), by combining the neighborhood knowledge granularity with the neighborhood discrimination index within the framework of neighborhood rough sets. NGDI can not only measure the uncertainty of feature subsets’ discriminant capability, but also characterize the granularity of neighborhood knowledge induced by feature subsets. Second, we propose an ensemble learning algorithm, ELNGDI, established on the NGDI. ELNGDI disturbs the feature space using multiple NGDI-based neighborhood approximate reducts. Third, we use ELNGDI to predict software defects. ELNGDI and the Synthetic Minority Oversampling Technique (SMOTE) are combined in order to handle the class imbalance issue in SDP, and propose a mechanism called SMOTE-ELNGDI. Experimental results on 20 datasets demonstrate that ELNGDI effectively improves the performance of SDP compared with existing ensemble learning methods.