Disposable Personas in Personalized Systems: Balancing Privacy and Usefulness

Abstract

Static user profiles help personalization systems make recommendations that are more relevant to each user. However, they can also reveal sensitive information about users by exposing patterns in their behavior. This paper introduces a framework, the disposable persona, to evaluate context-specific recommendations. The study compares a single persistent profile against context-specific personas using the DePaulMovies dataset. After deleting the avatar's history, the framework measures both how useful the recommendations are and how much context information is revealed as new ratings are added. Two main thresholds are used: k_useful, the number of ratings needed after a reset to regain useful personalization. Then k_private, the number of ratings after which context leakage becomes a concern. The framework tests four hypotheses about how this trade-off works. The results show that k_useful is always smaller than k_private for the three context dimensions tested.