Differentially Private Link Prediction with Protected Connections

Link prediction (LP) algorithms propose to each node a ranked list of nodes that are currently non-neighbors, as the most likely candidates for future linkage. Owing increasing concerns about privacy, users (nodes) may prefer keep some their connections protected or private. Motivated by this observation, our goal is design differentially private LP algorithm, which trades off between privacy node-pairs and link accuracy. More specifically, we first form differential on graphs, models loss only those marked protected. Next, develop DPLP, learning rank applies monotone transform base scores from non-private system, then adds noise. DPLP trained with induced ranking loss, optimizes utility given maximum allowed level leakage node-pairs. Under recently introduced latent embedding model, present formal trade-off utility. Extensive experiments several real-life graphs heuristics show can trade predictive performance more effectively than alternatives.