by Oxana Chakoula
Sequences of users' requests to a web server can be robustly modeled as Markov chains. There is a tradeoff between the order of a model (and its number of states) and predictive accuracy. We model web sessions as a probabilistic mixture of first order Markov chains. By grouping sessions of similarly-minded users and utilizing this knowledge for prediction, we aim to improve predictive power while retaining lower model complexity.