DIPF database of publications
Contextual models for user interaction on the Web
In: ECML-PKDD (Hrsg.): ECML/PKDD Workshop on Mining and Exploiting Interpretable Local Patterns (I-PAT)
URL of full text:
4. Beiträge in Sammelwerken; Tagungsband/Konferenzbeitrag/Proceedings
Accurately modeling user sessions on the web is important because such models can be used, on one hand to predict a user's actions, and on the other hand to inform design and content decisions. This includes predicting what links a user will click on, deciding where webpage components should be placed, and what content to provide. Often it is either undesirable or not possible to build personalized models, and even when available, such models suffer from the cold start problem, or are unable to deal with context-dependent variations in user behavior. In this paper, we present a probabilistic framework for session modeling that creates clusters of similar sessions and uses contextual session information (time, referrer domain, link locations). Sessions are probabilistically assigned to the clusters by conditioning on the context. The framework addresses wide variations in user behavior that are due to context by explicitly incorporating it in the model, while specifically leveraging periodicity (weekly and daily behavioral regularities). We evaluate the framework on a set of logs from Yahoo! News.