Anonymizing user profiles for personalized web search.

We study the problem of anonymizing user profiles so that user privacy is sufficiently protected while the anonymized profiles are still effective in enabling personalized web search. We propose a Bayes-optimal privacy notion to bound the prior and posterior probability of associating a user with an individual term in the anonymized user profile set. We also propose a novel bundling technique that clusters user profiles into groups by taking into account the semantic relationships between the terms while satisfying the privacy constraint. We evaluate our approach through a set of preliminary experiments using real data demonstrating its feasibility and effectiveness.