A new Markov model for Web access prediction

Abstract

T he World Wide Web is a large, distributed hypertext repository of information, which users navigate through links and view with browsers. The heavy Internet traffic resulting from the Web's popularity has significantly increased userperceived latency. The obvious solution-to increase the bandwidth-is not viable, because we cannot easily change the Web's infrastructure (the Internet) without significant economic cost. However, if we could predict future user requests, we could put those pages into the clientside cache when the browser is free. When a user requests one of the pages, the browser could retrieve it directly from cache.Much current research has examined modeling and predicting user access behavior on the Web to improve Web prefetching, 1,2 enhance search engines, 3 and understand and influence buying patterns. 4 To predict Web access, we need a method for modeling and analyzing Web access sequences. With this information, we can deduce future user requests.Some researchers have used traditional Markov models, which are often employed to study stochastic processes and predict user access behavior. 5,6 In general, they use the sequence of Web pages a user has accessed as input, with the goal of building Markov models with which they can predict the page the user will most likely access next. Venkata N. Padmanabhan and Jeffrey C. Mogul used N-hop Markov models to improve prefetching strategies for Web caches; 2 Ramesh R. Sarukkai used Markov models to predict the next page accessed by the user; 7 and Igor V. Cadez and colleagues used Markov models to categorize user sessions. 8 Peter Pirolli and colleagues, however, tested the performance of different-order Markov models for Web access prediction and found traditional Markov models to be inadequate for this purpose. 5 Therefore, we need a new Markov model for Web access prediction.Our hybrid-order tree-like Markov model can predict Web access precisely, providing high coverage and good scalability. HTMM intelligently merges two methods: a tree-like structure Markov model method that aggregates the access sequences by pattern matching and a hybrid-order method that combines varying-order Markov models. Performance evaluations comparing our HTMM with traditional Markov models confirm its usefulness.Accurately predicting Web user access behavior can minimize user-perceived latency, which is crucial in the rapidly growing World Wide Web. Although traditional Markov models have helped predict user access behavior, they have serious limitations. Hybrid-order treelike Markov models predict Web access precisely while providing high coverage and scalability.

A taxonomy of web prediction algorithms

“…Chen and Zhang [8] propose a memory-efficient version of a PPM algorithm whose prediction tree branches have different height depending on the root object popularity. A similar algorithm, using Markov models of different orders, is proposed by Dongshan and Junyi [20] to deal with the scalability of the prediction engine. The model presented by Gunduz and Ozsu [24] compare session similarities to make predictions.…”

Cost-Benefit Analysis of Web Prefetching Algorithms from the User’s Point of View

“…The experiments were run using three of the most widely used prediction algorithms in the literature: two main variants of the Prediction by Partial Match (PPM) algorithm [13,7,8,9] and the Dependency Graph (DG) based algorithm [2,9].…”

“…Predictions are obtained from the comparison of the current context to each Markov model. PPM algorithm has been proposed to be applied either to each object access [13] or to each page (i.e., to each container object) accessed by the user [7,8]. In this paper we implement the object-based version of the algorithm.…”

“…Some papers comparing the performance of prefetching algorithms have been published [7,8,9,10,11] but they mainly concentrate on predictive performance [7,8,9,10]. In addition, performance comparisons are rarely made using a useful costbenefit analysis, i.e., latency reduction as a function of the traffic increase.…”

“…In addition, performance comparisons are rarely made using a useful costbenefit analysis, i.e., latency reduction as a function of the traffic increase. As examples of some timid attempts, Dongshan and Junyi [7] compare the accuracy, the model-building time, and the prediction time in three versions of a predictor based in Markov chains. Another current work by Chen and Zhang [8] implements three variants of the PPM predictor by measuring the hit ratio and traffic under different assumptions.…”