Mining Web navigation patterns based on a mixture of latent variable models

The discovery and analysis of Web users' browsing patterns is important for gaining insight into their navigational behavior and information needs. Typical Web usage mining techniques, such as clustering of user sessions, or association rule discovery of Web page co-occurrence patterns, cannot capture the latent structures in the observation data. Nor can they quantify the hidden relationships between users' preferences and their observed behavior. This thesis proposes a unified framework to study the quantitative relationships between Web users' hidden preferences and their navigational behaviors from a probabilistic generative perspective. In particular, the framework is based on a mixture of latent variable models where continuous latent dimensions model users' hidden preferences and a discrete mixture variable models distinct user segments.; In addition to Web usage data, other knowledge sources such as the Web site content and linkage structure can be useful for interpreting, analyzing and reasoning about navigational patterns. The proposed framework results in a new multi-source integration model, extending the mixture of latent variable models, to seamlessly incorporate multiple knowledge sources and enrich the user models.; The extensive experimental results show that this framework is useful in uncovering the underlying user preferences and intentions behind their manifest browsing behavior as well as offering a great advantage in terms of probabilistic inference flexibility. Based on the framework, the thesis also demonstrate several useful applications in Web analytics and business intelligence such as Web user segmentation and automatic personalization.