Evaluating the performance and privacy of a token-based collaborative recommender

The rapid expansion of available online services has raised concerns about user privacy. In the online world, only a minority of users is actually aware where their data is stored and the policies, how the data may be eventually used. However, at the same time consumers expect more quality from online services, demanding personalized services that fit their individual needs, preferences and values. One approach for service personalization is to use collaborative recommenders. From the privacy perspective, mainstream collaborative recommenders present an inherent security risk, since they are based on memorizing user-item transactions. In this paper, we will study a recently developed token-based method (sometimes referred as an acronym "upcv") which creates privacy-protecting abstraction that is based on collections of randomly generated tokens. These collections are capable of providing information for collaborative recommendations without maintaining any transactional history. This paper presents quality evaluation of item-to-item recommendations using the token-based collaborative recommender, utilizing ISBN agencies of Book-Crossing dataset (BX) books at the data set. This paper will also discuss challenges related to BX. Privacy issues are evaluated with a specific emphasis on the concept of deniability.