Generative Frontier Planning for Adaptive Peer-Referral Recruitment under Covariate-Dependent Arrivals

Researchers propose Generative Frontier Planning (GFP), a novel algorithm for optimizing peer-referral recruitment in hidden populations by modeling realistic homophily effects and covariate-dependent arrivals. The method outperforms existing baselines by using deterministic backups over generative models rather than Monte-Carlo sampling, achieving near-optimal resource allocation for public health interventions.

This research addresses a critical gap in public health recruitment methodology by tackling the computational and modeling challenges of peer-referral systems like respondent-driven sampling. Traditional approaches assume homogeneous, i.i.d. referral patterns, which fails to capture the clustering and social context that characterize real recruitment networks. The authors introduce a realistic framework where referral capacity and recruit characteristics depend on individual referrers, fundamentally altering how planners should allocate resources across recruitment rounds.

The innovation lies in Generative Frontier Planning's architectural design, which replaces computationally expensive Monte-Carlo sampling with deterministic backups using learned generative models. By leveraging finite-dimensional summaries of the offspring distribution amortized offline, GFP reduces per-round planning complexity while maintaining theoretical guarantees. The algorithm achieves a (1-1/e)-approximation ratio for greedy allocation, a meaningful bound in optimization contexts. This mathematical structure enables tractable planning that scales practically.

For public health implementation, GFP offers tangible benefits in studying infectious disease patterns, drug use, and other stigmatized conditions where hidden populations resist traditional sampling. Optimized recruitment accelerates data collection, reduces costs, and improves intervention targeting. The simulation results demonstrate consistent outperformance against reinforcement learning and dynamic programming baselines, suggesting robustness across different discount scenarios. The framework's flexibility to incorporate data-driven covariate models makes it adaptable to diverse recruitment contexts and populations with varying homophily patterns.

• →Generative Frontier Planning improves peer-referral recruitment by modeling realistic homophily and covariate dependencies rather than assuming i.i.d. arrivals
• →The algorithm replaces Monte-Carlo sampling with deterministic backups over learned generative models, reducing computational complexity while maintaining approximation guarantees
• →GFP achieves (1-1/e)-approximation for greedy allocation by leveraging diminishing-returns structure in per-round objectives
• →Simulation results on calibrated respondent-driven sampling data show consistent outperformance over reinforcement learning and dynamic programming baselines
• →The approach has direct applications for public health agencies optimizing resource allocation in hidden population studies and disease interventions