Coupons meant to grow an e-commerce platform often just shuffle spending around — a paper from arXiv details a framework designed to fix that.
Researchers propose CanniUplift, a unified framework targeting two specific failure modes in personalized incentive allocation. The first is seller-level cannibalization: a discount drives a purchase, but from a competing shop on the same platform, leaving total gross merchandise value flat. The second is incentive-level cannibalization: a customer who would have bought anyway redeems a coupon, inflating the apparent lift from that promotion. To address both, the framework combines Platform-level Global Alignment (PGA), which enforces cross-shop GMV consistency constraints, and Redemption-based Decomposition Denoising (RDD), which uses redemption behavior to strip attribution noise from treated outcomes. A third component, a Treat-Attention mechanism, models interactions between a user's purchase history and the current incentive on offer.
The practical stakes are real: in large-scale e-commerce, even small misfires in uplift estimation mean budget allocated to customers who were already going to convert — or to deals that merely redirect spend without adding it. The authors report a 4.08% relative increase in incremental GMV over a production baseline in online A/B tests, alongside improved ROI, after deploying CanniUplift at scale.
Uplift modeling has been a standard tool in e-commerce promotion for years, but the SUTVA assumption it typically relies on — that one user's treatment does not affect another's outcome — breaks down badly in multi-seller marketplaces. CanniUplift's framing of that breakdown as two distinct, addressable problems is the more useful contribution here than any single performance number.