Seeding with Differentially Private Network Information
Authors:
M. Amin Rahimian,
Fang-Yi Yu,
Carlos Hurtado
Abstract:
When designing interventions in public health, development, and education, decision makers rely on social network data to target a small number of people, capitalizing on peer effects and social contagion to bring about the most welfare benefits to the population. Develo** new methods that are privacy-preserving for network data collection and targeted interventions is critical for designing sus…
▽ More
When designing interventions in public health, development, and education, decision makers rely on social network data to target a small number of people, capitalizing on peer effects and social contagion to bring about the most welfare benefits to the population. Develo** new methods that are privacy-preserving for network data collection and targeted interventions is critical for designing sustainable public health and development interventions on social networks. In a similar vein, social media platforms rely on network data and information from past diffusions to organize their ad campaign and improve the efficacy of targeted advertising. Ensuring that these network operations do not violate users' privacy is critical to the sustainability of social media platforms and their ad economies. We study privacy guarantees for influence maximization algorithms when the social network is unknown, and the inputs are samples of prior influence cascades that are collected at random. Building on recent results that address seeding with costly network information, our privacy-preserving algorithms introduce randomization in the collected data or the algorithm output, and can bound each node's (or group of nodes') privacy loss in deciding whether or not their data should be included in the algorithm input. We provide theoretical guarantees of the seeding performance with a limited sample size subject to differential privacy budgets in both central and local privacy regimes. Simulations on synthetic and empirical network datasets reveal the diminishing value of network information with decreasing privacy budget in both regimes.
△ Less
Submitted 1 February, 2024; v1 submitted 25 May, 2023;
originally announced May 2023.
