Abstract. As the foundation of network cognition, management and optimizing, the classification of network traffic is making a significant difference in resource scheduling, safety analysis and future tendency prediction. Feature subset selection (FSS) based on machine learning plays an important role in classification problems, especially dealing with high-dimensional data like network traffic flows. To realize accurate traffic classification at lower price of evaluations, a hybrid feature subset selection method is proposed on the base of dynamic block, the size of which is flexible according to the classification performance. The performances are examined a few experiments. Our theoretical analysis and experimental observations reveal that the proposed method consumes fewer evaluations with similar or even better classification performance.