Date Added: Oct 2009
In P2P VoD streaming systems, user behavior modeling is critical to help optimise user experience as well as system throughput. However, it still remains a challenging task due to the dynamic characteristics of user viewing behavior. In this paper, the authors consider the problem of user seeking prediction which is to predict the user's next seeking position so that the system can proactively make response. They present a novel method for solving this problem. In their method, frequent sequential patterns mining is first performed to extract abstract states which are not overlapped and cover the whole video file altogether. After mapping the raw training dataset to state transitions according to the abstract states, they use a simple probabilistic contingency table to build the prediction model.