Human Immunodeficiency Virus (HIV) remains a major global public‑health challenge. Deploying limited prevention and treatment resources efficiently requires precise network‑based interventions, yet the true contact network is costly to obtain at scale. Peer‑referral networks, collected through recursive invitation, can be gathered digitally from hidden populations but deviate from the underlying contact graph: referrals may follow non‑risk social ties and each participant accepts only one referral, obscuring links among already recruited individuals.
We first formalize the problem of planning interventions that minimize HIV spread in the hidden contact network while only observing the peer‑referral network. We then introduce R2C‑Agent, a decision‑making agent built around a novel conditional graph diffusion model called Referral‑to‑Contact Diffusion (R2C‑Diff). R2C‑Diff treats the peer‑referral process as incremental corruption of the underlying contact graph and learns a probabilistic reverse diffusion. R2C‑Agent samples contact graphs from R2C‑Diff and plans interventions on them, thereby accounting for uncertainty about the hidden network.
Simulated experiments on real‑world HIV network data compare R2C‑Agent against several interdisciplinary baselines. Results show that R2C‑Agent consistently yields more effective interventions. The study also examines how model assumptions, sample size, and other factors influence performance, highlighting key trade‑offs for practical deployment.
Review: By modeling observation error as a learnable diffusion process, R2C‑Agent offers an interpretable and scalable tool for resource‑constrained public‑health interventions, and its promise warrants validation in larger epidemiological settings.