Modeling Topic Selection of Web Browsing Using Clickstream Data

• Activity Number: 443
• Type: Invited
• Date/Time: Wednesday, August 3, 2011 : 8:30 AM to 10:20 AM
• Sponsor: IMS
• Author(s): Alan Montgomery*+
• Companies: Carnegie Mellon University
• Keywords: Bayesian Analysis ; Topic Models ; Clickstream Data ; Consumer Behavior ; Marketing

Abstract

Users appear to view websites in sequences that are related to a latent topic. For example, a user may have a session that is focused on gathering information about a product, which may have a large number of viewings at promotional, corporate, and portal sites, while news gathering may have a very different profile. The idea is that consumers group their activities together into related topics. The goal of this study is to detect the underlying topics that are driving user browsing behavior using a correlated topic model. Specifically we use a multivariate normal to model the log-odds ratio of a given topic being chosen. The correlations permit relationships amongst the topics. Conditional upon the topic each website is chosen using a choice model across all websites. Our model is related to a latent Dirichlet allocation and correlated topic models employed in text analysis. We consider generalizations with dynamic trends to understand how topic selections may depend upon time.