Infinite mixtures of Markov chains

Abstract

Facilitating a satisfying user experience requires a detailed understanding of user behavior and intentions. The key is to leverage observations of activities, usually the clicks performed on Web pages. A common approach is to transform user sessions into Markov chains and analyze them using mixture models. However, model selection and interpretability of the results are often limiting factors. As a remedy, we present a Bayesian nonparametric approach to group user sessions and devise behavioral patterns. Empirical results on a social network and an electronic text book show that our approach reliably identifies underlying behavioral patterns and proves more robust than baseline competitors.