Privacy-Preserving Link Prediction in Decentralized Online Social Networks

We consider the privacy-preserving link prediction problem in decentralized online social network (OSNs). We formulate the problem as a sparse logistic regression problem and solve it with a novel decentralized two-tier method using alternating direction method of multipliers (ADMM). This method enables end users to collaborate with their online service providers without jeopardizing their data privacy. The method also grants end users fine-grained privacy control to their personal data by supporting arbitrary public/private data split. Using real-world data, we show that our method enjoys various advantages including high prediction accuracy, balanced workload, and limited communication overhead. Additionally, we demonstrate that our method copes well with link reconstruction attack.