Learning to target with network interference

A new study posted to arXiv on 27 May 2026 establishes fundamental limits for adaptive targeting when treatments spill over across social or economic networks, showing that ignoring these connections guarantees inefficient learning [1][2]. The paper, "Learning to target with network interference," examines a bandit setting where a treatment applied to one individual can affect the outcomes of others through spillover effects [1][2]. The authors adopt a linear model in a sparse regime, meaning each person's outcome is influenced by at most a few others [1][2]. They first derive a regret lower bound demonstrating that treating the problem as a standard linear bandit — effectively discarding the network — leads to inefficient learning, especially as the population grows large [1][2]. To understand how structural information can improve performance, the researchers analyze three distinct regimes of prior knowledge about the interference structure: full support knowledge, knowledge of the column support sizes, and no prior knowledge [1][2]. For each regime, they establish regret lower bounds that characterize the fundamental limits of learning and develop algorithms that achieve near-optimal regret [1][2]. The work provides a unified view of how knowledge of the interference structure governs the efficiency of online learning under interference [1][2]. The concept of interference between units has parallels in other fields. In machine learning, catastrophic interference describes the tendency of an artificial neural network to abruptly forget previously learned information upon learning new data, a radical manifestation of the stability-plasticity dilemma [3]. While that phenomenon concerns sequential learning within a single model, the network interference studied here concerns cross-unit spillovers in a population. Neural networks themselves are computational models inspired by biological neural networks, consisting of connected units called artificial neurons that process signals through weighted connections [4]. Training such networks is a compute-intensive process accelerated by graphics processing units and large datasets [4]. In biological systems, motor learning research examines how the nervous system processes feedback to change connectivity and synaptic strengths, enabling organisms to acquire new skills and improve movement smoothness over time [5]. The timing and organization of practice can influence information retention, and the precise form of feedback can influence preparation, anticipation, and guidance of movement [5]. The new arXiv paper extends the study of learning under structured interference into the algorithmic domain, with numerical experiments on synthetic and real-world data demonstrating practical benefits [1][2]. The preprint was submitted through arXivLabs, a framework that allows collaborators to develop and share new features on the arXiv platform [1].