Learning to Allocate Propagating Treatments: Maximize Uplift under Network Interference

Ruiheng Jiang, Xinyan Su, Han Huang, Zhiheng Zhang

Keywords: causality; network interfernece

TL;DR: Learning to Allocate Propagating Treatments: Maximize Uplift under Network Interference

Abstract: Uplift modeling and treatment allocation are classical tasks in promotion marketing. Yet existing allocations ignore propagating treatments and network interference, where both outcomes and the propagation mechanism vary with peers’ treatment history, making policy value hard to estimate and optimize. We formalize a history-driven uplift objective with activation probability $g(\bm Z_i^t,\bm X_i)$ and outcomes that depend on neighbors’ treated states. Theoretically, we establish conditions for identification and provide finite-sample guarantees for policy evaluation under interference and model misspecification. Methodologically, we propose GUM-DT via a Monte-Carlo policy search: learn an ensemble of lightweight propagation models and an outcome model, and evaluate candidate allocations via double-robust (DR) estimators with IPW corrections. On synthetic networks, experiments demonstrate consistent gains over uplift allocations of GUM-DT, validating robustness and effectiveness.

Primary Area: causal reasoning

Submission Number: 3769