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

1 Aggregation Page Ads

Aggregation page ads are a key scenario in Meituan’s advertising business, displaying merchants and multiple coupons in a combined view. The article focuses on two decision‑critical aspects—coupon issuance and ranking—and explores context‑aware modeling to improve personalization.

2 Practice 1: Context‑Enhanced Full‑Chain Uplift Modeling (ECUP)

2.1 Background

In e‑commerce marketing, treatment design (e.g., coupons) and effect evaluation are essential. Uplift models predict the individual treatment effect (ITE) to allocate coupons without waste. Traditional uplift methods treat each stage (exposure → click → conversion) independently, leading to chain bias and treatment‑adaptation issues.

2.2 Chain‑Bias Analysis

Using the public Criteo‑Uplift dataset and Meituan’s MT‑LIFT dataset, the authors compute uplift for CTCVR (exposure space) and CVR (click space) across random segments. The trends differ, showing that click‑based CVR uplift does not reliably reflect full‑chain uplift, because user focus varies across stages.

2.3 Method

ECUP consists of two networks:

Full‑Chain Enhancement Network (ECENet) estimates results for every task in the chain and injects task priors via a Task‑Enhanced Network (TAENet).

Treatment Enhancement Network (TENet) refines initial features with treatment‑aware embeddings at the bit level, addressing treatment‑adaptation.

TAENet uses a gated mechanism that treats task information as a query and treatment embeddings as key/value in a multi‑head attention module, allowing dynamic parameter selection per layer. TENet comprises a Treatment‑Aware Unit (TAU) and a Treatment‑Enhanced Gate (TEGate) that perform bit‑level weighting of treatment‑aware features.

2.4 Effectiveness

The authors collected two months of coupon‑marketing data from the aggregation page, forming a dataset of ~5.5 M instances, 99 features, and two labels (click, conversion). Offline experiments on public and Meituan datasets show superior AUUC and QINI scores (see

), and ablation studies confirm the contribution of each module (see

). An online A/B test reported statistically significant lift (see

).

3 Practice 2: Neighbor‑List Guided Rerank (NLGR)

3.1 Background

Reranking reshuffles an initial list to improve multi‑stage recommendation. Existing generator‑evaluator frameworks suffer from objective mismatch and ignore downstream items, especially when modules such as “one‑click coupon” and “regular coupon” have divergent CTR patterns.

3.2 Method

NLGR follows the generator‑evaluator paradigm but enhances the generator with neighbor‑list information. During training, the generator replaces items at selected positions (PDU) and inserts candidates (CDU) drawn from neighboring lists, using Gumbel‑Softmax for differentiable sampling. The evaluator remains a standard MLP with cross‑entropy loss. The training process constructs neighbor lists and uses their rewards to guide both PDU and generator updates (see

).

3.3 Effectiveness

Offline results on public and Meituan datasets show higher AUC for the evaluator and higher hit‑ratio for the generator (see

and

). Online A/B experiments on the external aggregation page demonstrate notable revenue gains (see

).

4 Summary and Outlook

Context‑aware modeling of aggregation page ads yields significant improvements in both coupon issuance and ranking. Future work will explore more efficient counterfactual bias correction, dynamic budget modeling for coupons, globally aware generative‑evaluator methods, and the integration of large language models with full‑chain context.