Web User Profiling on Proxy Logs and Its Evaluation in Personalization

摘要

We propose a web user profiling and clustering framework based on LDA-based topic modeling with an analogy to document analysis in which documents and words represent users and their actions. The main technical challenge addressed here is how to symbolize web access actions, by words, that are monitored through a web proxy. We develop a hierarchical URL dictionary generated from Yahoo! Directory and a cross-hierarchical matching method that provides the function of automatic abstraction. We apply the proposed framework to 7500 students in Osaka University. The framework is used to analyze their 40GB click streams over a 4 month period. We evaluate clustering-based recommendation effectiveness to confirm the optimality of the framework. The results show high hit precision compared with existing methods.