An online PPM prediction model for web prefetching

Abstract

Web prefetching is a primary means to reduce user access latency. An important amount of work can be found by the use of PPM (Prediction by Partial Match) for modeling and predicting user request patterns in the open literature. However, in general, existing PPM models are constructed off-line. It is highly desirable to perform the online update of the PPM model incrementally because user request patterns may change over time. We present an online PPM model to capture the changing patterns and fit the memory. This model is implemented based on a noncompact suffix tree. Our model only keeps the most recent W requests using a sliding window. To further improve the prefetching performance, we make use of maximum entropy principle to model for the outgoing probability distributions of nodes. Our prediction model combines entropy, prediction accuracy rate and the longest match rule. A performance evaluation is presented using real web logs. Trace-driven simulation results show our PPM prediction model can provide significant improvements over previously proposed models.