The Navigation Problem in the World-Wide-Web

Herein we build statistical foundations for tackling the navigation problem users encounter during web interaction, based on a formal model of the web in terms of a probabilistic automaton, which can also be viewed as a finite ergodic Markov chain. In our model of the web the probabilities attached to state transitions have two interpretations, namely, they can denote the proportion of times a user followed a link, and alternatively they can denote the expected utility of following a link. Using this approach we have developed two techniques for constructing a web view based on the two interpretations of the probabilities of links, where a web view is a collection of relevant trails. The first method we describe is concerned with finding frequent user behaviour patterns. A collection of trails is taken as input and an ergodic Markov chain is produced as output with the probabilities of transitions corresponding to the frequency the user traversed the associated links. The second method we describe is a reinforcement learning algorithm that attaches higher probabilities to links whose expected trail relevance is higher. The user’s home page and a query are taken as input and an ergodic Markov chain is produced as output with the probabilities of transitions giving the expected utility of following their associated links. Finally, we characterise typical user navigation sessions in terms of the entropy of the underlying ergodic Markov chain.