Learning from What Others Know: Privacy Preserving Cross System Personalization

Recommender systems have been steadily gaining popularity and has been deployed by several service providers. Large scalable deployment has however highlighted one of the design problems of recommender systems: lack of interoperability. Users today often use multiple electronic systems offering recommendations, which cannot learn from one another. The result is that the end user has to often provide similar information and in some cases disjoint information. Intuitively, it seems that much can be improved with this situation: information learnt by one system could potentially be reused by another, and new insights learnt by a third system about the same user can be propagated to other systems, to offer an overall improved personalization experience. In this paper, we provide an effective solution to this problem using Latent Semantic Models by learning a user model across multiple systems. A privacy preserving distributed framework is added around the traditional Probabilistic Latent Semantic Analysis framework, and practical aspects such as addition of new systems and items are also dealt with in this work.