Data Mining of Digital Behaviour

点击翻译按钮获取中文摘要

The goal of this research project is to improve our understanding of how humans behave in information-seeking digital environments such as the Web. The approach consists of using massively large Web logs to infer patterns of behavior. New probabilistic models for modeling human behavior on the Web are under investigation including Markov and switching models, mixture models, and Bayesian hierarchical models. Adaptive statistical techniques form the basis for building up individual user models in an online fashion, automatically learning both the dynamic time-dependent patterns of a user as well as text-vector representations of their interests. Test data sets from large commercial Web sites are being used to develop, validate, and test the models. Data are anonymized to protect individual privacy. The statistical user models are in turn being used to develop two primary software tools. The first tool allows an analyst to explore, cluster, predict, and visualize Web logs with millions of entries, allowing an understanding of dynamic patterns of access and behavior in a manner that is not currently available in research or commercial tools. The second tool, WebMARS, uses adaptive user-models to enhance information retrieval algorithms by interpreting search queries in a personalized manner. More generally, the results from this project will provide tools and techniques to enable a better scientific understanding of modes of human behavior across a broad range of digital environments, with potential applications in wireless information appliances, medical informatics, scientific exploration of massive data sets, and so forth.