A PPM Prediction Model Based on Stochastic Gradient Descent for Web Prefetching

Abstract

PPM models are commonly used to predict the user's next request in Web prefetching by extracting useful knowledge from historical user requests. Any of features such as page access frequency, prediction feedback, context length and conditional probability can influence on the prefetching performance of PPM models. However, existing PPM models only consider one or a few of them. Based on stochastic gradient descent, we present a novel PPM model that takes into account all the above mentioned features. Our model defines a target function to describe a node's prediction capability, which is a linear combination of these features. In order to decrease the number of incorrect predictions, weights of these features are dynamically updated over every example according to the stochastic gradient descent rule. Our model selects a node with the maximum target function value among all matching nodes to predict the next most probable page. We use real web logs to examine proposed model. The simulation shows our model can significantly improve the prefetching performance.