Oversampling Algorithm based on Reinforcement Learning in Imbalanced Problems

The imbalanced problem indicates that the data set is unevenly distributed, resulting in sub-optimal classifiers to recognize the minority class. Traditional solutions try to design new classifiers to solve this problem or balance the skewed data sets, the former is too costly while the latter has an uncertain effect on different combinations of classifiers and measurements. In this paper, we propose a reinforcement learning-based oversampling method, which can directly produce targeted samples according to the downstream classifiers and measurements. During training, our learning procedure introduces the classification information to the generation process. Moreover, as opposed to oversampling approaches, we have no assumption of the downstream classifiers and performance metrics, and the proposed has a wider application. We carry out experiments on 17 UCI and KEEL data sets, experimental results demonstrate the superior performance of our proposed method.