News
KDD ADS submission on multi-channel uplift policy learning
A KDD 2027 ADS submission introduces ReAlloc for support-aware causal reallocation across multiple marketing channels.
We submitted Multi-channel Uplift Policy Learning to the KDD 2027 Applied Data Science (ADS) Track.
The paper studies how an e-commerce platform should allocate a fixed marketing budget across multiple channels when historical decisions are confounded and global optimization can extrapolate beyond observed support.
It introduces ReAlloc, a fast-slow causal framework in which an orthogonal teacher estimates local response gradients from recent logs, an explanation-guided student distills them into a structured marginal field, and a support-aware policy performs conservative reallocations. Simulations, offline production-log evaluation, and large-scale Taobao A/B tests show simultaneous gains in pay order and income.