Abstract: In the process of data stream classification, class imbalance and concept drift are two major challenges. On the basis of feature selection and cost-sensitive learning, the idea of cost-sensitive learning algorithm is introduced into the feature selection algorithm, and a cost-sensitive ReliefF method is designed and implemented. The algorithm can not only delete redundant features, but also adapt to dynamic data stream environment. The classical algorithms are analyzed and compared, and the results show that the proposed algorithm can significantly improve the performance of classification.