Targeted AI Deployment in Expert Decision-Making: A Dyadic Policy Learning Framework

数字平台 increasingly use AI to augment, not replace, human experts, creating

a need to target support based on heterogeneous expert–task interactions.

Existing methods, which rely on i.i.d. assumptions and plug-in rules,

often ignore first-stage estimation error and fail to quantify policy

uncertainty. We introduce a dyadic policy learning framework that combines

doubly robust estimation with two-way cross-fitting to address these limitations.

Establishing uniform regret bounds under separate exchangeability, our

approach uses a nested diagonal cross-fitting design and pigeonhole bootstrap

for honest welfare evaluation. Applied to radiologists interpreting chest

X-rays, our learned tree-based policies yield statistically significant

welfare gains over no-AI baselines across diagnostic utility, ranking

quality, and efficiency. By adapting policy learning to dyadic structures,

we prevent the overfitting and biased precision estimates common in conventional

i.i.d. methods, offering a robust blueprint for embedding AI in high-stakes,

interdependent workflows.

张成龙 副教授 信息管理与商业智能系