This paper presents an effective method for solving imbalanced learning. Imbalanced learning is a recognition problem with data with imbalanced distributions. Many practical applications, e.g., fraud identification and intrusion detection, face the critical problem of imbalanced data. Most traditional methods use accuracy to evaluate the classifier's performance such that it is difficult to improve the accuracy of minority classes. In order to address the problem of imbalanced data, this proposed methods has two advantages. First, a new strategy of weight updating is proposed, called G-weighting scheme. It assigns different weights to different kinds of training samples such that these weights can represent the importance of training samples and the case that the classifier favors majority classes can be avoided. Secondly, a cascading framework is proposed to improve accuracy of each class and get further balanced performance between accuracy and G-means measure. In addition, this proposed method has the advantages of fast learning speed and good generation. Experimental results on 15 binary-class data and 5 multi-class data from KEEL repository show that this proposed method can get better results than other state-of-the-art methods.