ECUP and NLGR: Context-Aware Uplift Modeling and Reranking for Meituan Aggregation Page Ads

Meituan's aggregation page advertising displays merchants and coupons to users, with coupon issuance and ranking being the two most influential factors for user decisions. The paper studies context-aware modeling to improve these directions.

For coupon issuance, the authors propose ECUP (Entire Chain Uplift Modeling with Context-Enhanced Learning), which tackles chain bias and treatment mismatch. ECUP consists of a full-chain enhancement network (ECENet) that models each task's outcome using user sequences, a task-enhanced network (TAENet) that injects task priors, and a treatment enhancement network (TENet) that adapts embeddings at the bit level. Experiments on public and internal datasets show improved AUUC and QINI gains, validated by online A/B tests.

For ranking, the authors propose NLGR (Neighbor List Generative Rerank), an evaluator-generator framework where the generator is trained using neighbor lists to sense relative scores and a non-autoregressive sampling method enables flexible jumps. The evaluator uses standard cross-entropy loss, while the generator's reward comes from comparing neighbor list and candidate list rewards. NLGR improves hit ratio on public and Meituan datasets and yields significant benefits in online A/B experiments.

Overall, the work demonstrates the value of context-aware uplift modeling and neighbor-list-based reranking for aggregation page ads, and outlines future directions such as debiased counterfactual methods, budget-aware modeling, global generative/evaluative approaches, and LLM integration.