Abstract: Composite service selection is one of the core research issues in the service computing field. There are complex correlations among atomic service QoS in a composite service, making some atomic services become more efficient and some become less efficient if they are used together. Previous service selection algorithms almost ignored this problem, which makes the Web service QoS used to service selection be inaccurate and the selected composite service be not the optimal one in the real executing environment. To address this problem, a division based composite service selection approach is proposed in this paper. Firstly, the set of efficient composite service execution instances are extracted from the log repository. Then the concrete service sets which are frequently used together are mined. On this basis, the composite service is divided into some dots, and the corresponding dot patterns of each dot are generated. Finally, composite service selection can be done taking the divided composite service dots as selection units, and the dot patterns of each dot as its candidate service set. Because dot patterns are testified by many execution instances and represent selection experience, they are usually more efficient compared with the results of doing selection independently for each service in one dot. Experimental results show that the proposed approach can improve the quality of selected composite services effectively.