On the Compression of Markov Prediction Model

Markov prediction model is the basis of Web prefetching and personalized recommendation. The existence of a large amount of Web objects results in a vast increase in the number of states which represent the users visited transfer behavior, which also causes the problem of huge spatial complexity in prediction model. In view of the transition probability matrix in Markov prediction model, this paper proposes a measurement method based on row similarity and column similarity. First, the similarity matrix is calculated. Then the row similarity and column similarity are used to obtain similar pages simultaneously which can be compressed together. Thus the number of states can be reduced. The experimental results show that the model can not only have good overall performance and compression effect but also keeps relative higher prediction accuracy and recall.