Feature Generation Model for Imbalanced Classification Problems

Class-imbalance learning deals with the problem where the number of training data in one class is significantly more than another class. To tackle this problem, some feasible and effective methods have been proposed in recent years. However, most of them are simply designed and cannot solve the essential difficulties embedded in the imbalanced data classification like noises and class overlapping. These problems make learning imbalanced data more difficult, but existing algorithms tend to ignore it and directly utilize intuitive class-wise resampling or reweighting to balance the dataset. To address these challenges, we propose a feature generation model, which mainly contains two parts: competitive learning and feature transformation. The experiments result on eight class-imbalance datasets show that the proposed method outperforms state-of-the-art approaches in class-imbalance learning task.