Classifier learning with imbalanced data is one of the main challenges in the data mining community. An ensemble of classifiers is a popular solution to this problem, and it has acquired significant attention owing to its better performance as compared to individual classifiers. In this paper, we propose an imbalanced classification ensemble method, which is hereafter referred to as overlap and imbalanced sensitive random forest (OIS-RF). We consider the existence of overlap in imbalanced data and create a new coefficient called Hard To Learn (HTL) which aims to measure the degree of importance for each training instance. In this regard, OIS-RF focuses more on learning the instances with high importance in each sub-dataset. Furthermore, to encourage the diversity of the ensemble, a weighted bootstrap method is proposed to generate sub-datasets containing diverse local information. The proposed method is evaluated on imbalanced datasets and is supported by statistical analyses. The results show that our method outperforms 9 state-of-the-art ensemble algorithms.