In view of the problem that existing fake comments detection methods did not take full advantage of dynamic information contained in the history of user behavior, in the paper, we firstly mine the dynamic features from the information of user dynamic behavior by using the time series analysis model. Secondly, we integrate the dynamic features and the user-level static features to find the suspicious users, then the user suspicious probability is propagated to the user comments. Finally, we get the fake comment classification features by fusing comment suspicious probability and the comment-level static features, and then use the PU-Learning classification strategy to accomplish the detection of the fake comments. Experimental results show that, the method we proposed can effectively improves the accuracy of fake comments detection system.