HIV interventions no longer need the full contact network: R2C-Agent plans from peer referrals alone, cutting total regret 54.9% below the best baseline
Synopsis
HIV contact networks are prohibitively expensive to collect at scale, while peer referral networks can be gathered digitally but deviate from the true contact structure because referrals follow non-risk social ties and each participant accepts only one referral. R2C-Agent treats the referral process as incremental corruption of the contact network and learns to reverse it, then plans interventions over sampled contact networks. On real ICPSR HIV network data, its total regret is 54.9% lower on intensive treatment and 28.8% lower on targeted PrEP than the best baselines.
Interpretation
Intervention planning can now rely only on a peer referral network that is feasible to collect at scale, without first observing the full HIV contact network. Earlier network intervention algorithms generally assume the contact network is observable, or fall back on heuristics such as highest-degree selection and individual risk scores that ignore network effects. On real ICPSR HIV network data, measured as total regret against the unobserved contact network, R2C-Agent is 54.9% lower than the best baseline, Highest Degree, on intensive treatment and 28.8% lower than the best baseline, Identity, on targeted PrEP.
R2C-Agent treats each referral step as a known corruption of the clean contact network and reverses the steps one at a time to recover plausible contact networks. This conditional graph diffusion model, Referral-to-Contact Diffusion (R2C-Diff), builds the referral mechanism and the nudge order directly into the forward process instead of making the model learn from scratch how referral graphs relate to contact graphs. The agent samples several contact networks from R2C-Diff and passes them to the CAST and SandIMIN intervention algorithms for intensive treatment and targeted PrEP respectively, then evaluates decisions on the true contact network with 100 Monte Carlo spread samples.
Decision quality and network reconstruction accuracy do not move together, and sampling from the distribution suits intervention choice better than committing to a single most likely network. R2C-Diff's average edit distance is 1721.6, a factor of 1.33 worse than Identity at 1290.2, which uses the referral network directly, yet its interventions come out better. On intensive treatment, fixing the Monte Carlo sample budget at 50 while reducing the number of sampled contact networks and drawing more spread samples per network moves total regret from 53.8 to 101.3 and then to 240.4, indicating that exploring structural uncertainty matters more than exploring spread-outcome uncertainty.
Perspective
The evaluation rests on one anonymized, publicly available real HIV network from ICPSR with 1386 nodes, 248 people living with HIV and 3793 edges, of which 1381 are contact edges; referral networks are generated by a simulated recursive referral process, and intervention effects are measured by Monte Carlo spread simulation under the Independent Cascade model rather than a real intervention trial.
The referral simulation assumes nodes recruit along all of their contact and social edges, and an ablation adds synthetic edges to the true test network to probe violations of that assumption; training uses teacher forcing, so errors may accumulate at inference; the authors propose future decision-focused variants of R2C-Diff and ask whether the model generalizes to unseen networks when training data are so limited.