It is difficult for humans to behave randomly. Over the past decades, psychological research shows that we cannot randomize our responses, even if we specially try. Therefore, predicting a player's actions through opponent modelling is essential in design of game AI. For Online games, learning to anticipate one player's actions also is a great benefit to play against a completely different player. In this paper, we propose a novel N-Gram based algorithm and use a suffix matching tree to model opponent behaviours. By the model, we can predict the player's next style efficiently using the last few actions seen. Compared with our previous data stream mining approaches, naming MSSBE and MSSMB, an approximate 40% improvement is achieved in accuracy of prediction. The proposed method is simple in its implementation and has the potential for parallel processing in GPU.