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

Problem Formulation

The paper addresses the challenge of adaptive multi-round resource allocation in peer-referral recruitment systems, such as Respondent-Driven Sampling (RDS), used to study hidden populations affected by infectious diseases. Public health agencies possess a fixed total budget of referral resources (e.g., vouchers) to distribute across multiple rounds.

The core difficulty lies in the covariate-dependent nature of arrivals. Unlike prior work that assumes new recruits are drawn independently and identically (i.i.d.) from a homogeneous population, this work models a realistic scenario where:

1. Referral Capacity is Stochastic and Heterogeneous: The number of peers an individual can refer depends on their specific covariates (demographics, behavior, geography) and is subject to a censored count distribution.
2. Offspring Covariates are Conditional: The characteristics of new recruits are not drawn from a fixed distribution but are generated conditionally based on the referrer's covariates (reflecting homophily and shared context).

The goal is to design a policy that allocates resources to the current "frontier" of active referrers to maximize the expected discounted number of successful recruits over a finite horizon. This requires reasoning about how current allocation decisions shape both the size and the covariate composition of future frontiers.

Methodology: Generative Frontier Planning (GFP)

The authors propose Generative Frontier Planning (GFP), a model-based approximate dynamic programming method that overcomes the intractability of planning under covariate-dependent generative arrivals.

1. Learned Dynamics

GFP utilizes two learned components estimated offline from historical data:

• Censored Count Model ($q_\psi$): A neural Poisson regression model estimating the referral capacity $C_i$ of an individual given their covariates $x_i$. It accounts for the fact that observed referrals are capped by the allocated budget.
• Conditional Diffusion Model ($G_\theta$): A conditional diffusion model that generates the covariates of new recruits ($x'$) given the parent's covariates ($x$).

2. The Computational Bottleneck

Standard model-based planning (e.g., Fitted Value Iteration) is intractable here because:

• The next frontier is a random set of variable size.
• Different allocations induce different offspring distributions, preventing the sharing of Monte-Carlo samples across candidate actions.
• The action space (integer allocations across all frontier members) is combinatorially large.

3. Structured Value Surrogate

To resolve these issues, GFP introduces a structured value surrogate $V_\phi(r, F)$ that maps the frontier to a latent covariate-coverage representation.

• Latent Representation: Each individual is mapped to a non-negative latent vector $h_\phi(x) \in \mathbb{R}^d_+$, representing a soft assignment to $d$ "covariate prototypes." The frontier is summarized by the aggregate coverage vector $z = \sum_{x \in F} h_\phi(x)$.
• Value Function: The value is modeled as an exponential-saturation function over this coverage: $V_\phi(r, z) = \sum w_{\phi,j}(r)(1 - \exp(-z_j))$. This structure captures diminishing returns as specific covariate regions become saturated.

4. Closed-Form Deterministic Backup

The key innovation is the use of Conditional Laplace Embeddings to make the Bellman backup deterministic and closed-form.

• Instead of sampling offspring from the diffusion model $G_\theta$ during planning, GFP pre-computes (amortizes) embeddings $\alpha_{\theta,\phi,j}(x) = \mathbb{E}_{y \sim G_\theta(\cdot|x)}[\exp(-h_{\phi,j}(y))]$.
• Using these embeddings and the censored count model, the expected future value can be computed exactly without Monte-Carlo sampling. The expectation factorizes into a product of terms dependent only on the allocation $k$ and the pre-computed embeddings.

5. Intra-Round Greedy Allocation

With the deterministic surrogate, the per-round optimization problem becomes maximizing a function that is monotone with integer diminishing returns.

• GFP employs a marginal greedy algorithm to allocate resources within a round.
• Theoretical analysis proves that this greedy approach achieves a $(1 - 1/e)$-approximation of the optimal fixed-budget allocation for the surrogate objective.

Key Contributions

1. Problem Generalization: The authors formalize peer-referral recruitment as an adaptive allocation problem where future arrivals are generated conditionally on referrer covariates, moving beyond the restrictive i.i.d. population assumption of prior work.
2. Generative Frontier Planning (GFP): A novel planner combining learned censored count and conditional diffusion models with a latent-coverage value surrogate. It replaces per-step Monte-Carlo sampling with a deterministic backup using amortized conditional Laplace embeddings.
3. Theoretical Guarantees: The paper proves that the induced per-round objective satisfies integer diminishing returns, guaranteeing a $(1 - 1/e)$-approximation for the intra-round greedy allocation.
4. Empirical Performance: Extensive experiments on a simulation environment calibrated to a real-world RDS dataset (ICPSR 22140) demonstrate that GFP outperforms random allocation, reinforcement learning baselines (Budget-DQN, Factorized RL), and an i.i.d. population-level dynamic programming baseline across four discount factors.

Experimental Results

The evaluation was conducted in a simulation environment mirroring real-world RDS dynamics, using a total budget of 100 and an initial frontier of 10 individuals.

• Performance: GFP consistently achieved the highest final cumulative recruits and final discounted cumulative reward across all discount factors ($\gamma \in \{0.9, 0.95, 0.99, 1.0\}$).
• Comparison to Baselines:
  • vs. IID-Population DP: GFP significantly outperformed the strongest baseline (IID-Population DP), which uses dynamic programming but ignores covariate composition. At $\gamma=1.0$, GFP recruited 99.5 individuals compared to 91.3 for the baseline, isolating the value of reasoning about covariate composition.
  • vs. RL: Reinforcement learning baselines struggled due to the large combinatorial action space and lack of structural exploitation, performing significantly worse than model-based approaches.
• Dynamics: While some baselines recruited faster in early rounds, GFP's covariate-aware strategy allowed it to pace allocations effectively, overtaking others in later rounds to maximize total reach.

Significance and Claims

The paper claims that GFP provides a tractable solution to a previously intractable planning problem in public health recruitment. By explicitly modeling the dependency between referrers and recruits, the method enables agencies to allocate limited resources more effectively than strategies that assume a homogeneous population.

The authors emphasize that their approach is modest and data-driven:

• It relies on learned surrogates rather than perfect knowledge of the true dynamics.
• The theoretical guarantees apply to the surrogate objective, with the gap to the true optimal policy decomposed into model estimation errors and approximation errors.
• The work is validated on a simulation calibrated to real data, but the authors acknowledge that validation on actual recruitment data in partnership with public health agencies is a necessary next step.

Ultimately, the paper demonstrates that incorporating generative models of covariate-dependent arrivals into dynamic programming frameworks can significantly extend the reach of peer-referral systems, offering a practical tool for intervening in hidden populations.