UBC Theses and Dissertations
Predicting users’ next access to the web server Chakoula, Oxana
The World Wide Web is experiencing a rapid growth both in the volume of traffic and complexity of web sites. Two of the main directions of research on improving users' web browsing experience are reduction of network latency perceived by users, and website personalization. Both tasks require the development of models that can predict a user's next request to a web server. Markov models trained on web logs have been found well suited for addressing this problem. Most of the variations of Markov models suggested for predicting a user's next request to a web server base their predictions on the data from the entire user population. Our hypothesis is that clustering users' web sessions to reflect browsing patterns of particular groups of users will improve prediction accuracy. We consider two clustering techniques that use different representations of web sessions. One approach treats web sessions as Markov models of order zero, and the other represents them as Markov models of order one. We compare the two clustered prediction models to the state-of-the-art non-clustered Markov models of orders 0, 1, All-2 and All-3. We report empirical results based on web logs of UBC Computer Science department. We found clustering web sessions as Markov chains performed better than clustering them of Markov models of order zero, yet the former did not improve predictability over the non-clustered first order Markov model.
Item Citations and Data