Abstract: Based on concept transfer graph model, a data stream classification algorithm is proposed and designed to deal with mixed concept drift environment. The algorithm measures the concept drift by conceptual representation of the instance set on the data block, detecting the degree of change of the concept, and introduces a dynamic and adaptive threshold. The base classifier model is reasonably selected for each instance to be classified, making full use of the potential diversity of the base classifier model and reducing the classification error during drift recovery. The performance of the proposed algorithm is superior to other algorithms on most of data sets, and has good adaptability in complex concept drift environments.