Modeling Category Viewership of Web Users with Multivariate Count Models

We develop a statistical model of browsing behavior by predicting the number of web pages, in a particular category, that are viewed by a user in a single web session. The purpose of this analysis is to better understand web browsing behavior, and to help predict which sessions are likely to result in retail visits. A single record in our database consists of the number of web pages viewed by a user during a single session from each of the following categories: portals, services, entertainment, retail, auctions, adult, and others. This dataset can be characterized as multivariate count data, where many of the counts are zero. We consider the use of Poisson and discretized tobit models, and contrast both univariate and multivariate versions of these models. Additionally, as our dataset is characterized by a great deal of heterogeneity in usage across users and also a good deal of persistence in viewership, we propose a new multivariate tobit model with a mixture process whose multiple states are governed by an unobserved (hidden) Markov chain. We find that users move between sessions that are characterized by browsing behavior that is focused in specific categories and sessions characterized by a variety of categories being viewed.