Learning to Un-Rank

Abstract

Search engines in online communities such as Twi er or Facebook not only return matching posts, but also provide links to the pro les of the authors. us, when a user appears in the top-k results for a sensitive keyword query, she becomes widely exposed in a sensitive context. e e ects of such exposure can result in a serious privacy violation, ranging from embarrassment all the way to becoming a victim of organizational discrimination.In this paper, we propose the rst model for quantifying search exposure on the service provider side, casting it into a reverse knearest-neighbor problem. Moreover, since a single user can be exposed by a large number of queries, we also devise a learningto-rank method for identifying the most critical queries and thus making the warnings user-friendly. We develop e cient algorithms, and present experiments with a large number of user pro les from Twi er that demonstrate the practical viability and e ectiveness of our framework.