A Compression Framework for Generating User Profiles
Xiaoxiao Shi, Kevin Chen–Chuan Chang, Vijay K. Narayanan, Vanja Josifovski, Alexander Johannes Smola · 2010
Predicting user preferences is a core task in many online applications from ad targeting to content recommendation. Many prediction methods rely on the being able to represent the user by a profile of features. In this paper we propose a mechanism for generating such profiles by extracting features that summarize their past online behavior. The method relies on finding a compressed representation of the behavior by selecting the dominant features contributing to the Kullback-Leibler divergence between the default distribution over user actions and the user specific properties. We show that the feature selection model of [1] can be extended to a hierarchical encoding of user behavior by means of using an intermediate clustering representation. Preliminary experiments suggest the efficacy of our method.