Heterogeneous Ad Mixing and Candidate Expansion for Location-Based Services in Meituan In‑Store Advertising

Under Location‑Based Services (LBS) constraints, the limited number of store candidates restricts the potential of Meituan's in‑store advertising ranking system. This article introduces a candidate‑expansion approach from the perspective of candidate types and presents a high‑performance heterogeneous mixing network to overcome performance challenges, thereby raising the upper bound of the local‑life ranking system.

Meituan's in‑store advertising drives commercial monetization for search traffic, serving restaurants, entertainment, beauty, hotels, and other local services. The quality‑estimation team has previously published work on CTR/CVR and other quality scores at conferences such as SIGIR, ICDE, and CIKM. Those works mainly improved model accuracy, but ranking quality also depends on the candidate pool. Because LBS limits the number of store candidates for many categories, the overall ranking potential is severely constrained.

To address this, the system expands candidates by supplementing store ads with product ads. Two typical heterogeneous mixing patterns are described:

Competitive heterogeneous ads : store and product ads compete for the same slot; the ad with higher predicted pCTR wins. (See Figure 1)

Combinational heterogeneous ads : a store ad is displayed together with its top‑2 product ads as a single unit; the unit’s score is computed from the store and product predictions. (See Figure 2)

Because the original DNN model only scores about 150 store candidates, it lacks product‑level information. By extending the scoring granularity to the product level, the candidate count can increase to 1500 + , and with a generative combination estimator it can reach 1500× under online performance constraints. This expansion raises three major challenges:

Performance pressure at product granularity (10× more candidates).

Difficulty modeling the relationship between a store and its multiple products.

Cold‑start problem for newly introduced product ads.

The solutions are:

High‑performance heterogeneous mixing system : a shared‑representation (bias) network learns store features once and reuses them for all products, reducing duplicated computation. During offline training, a “store network” processes store features, while a “bias network” processes product features; their outputs are concatenated and jointly optimized. At inference, a list of products is fed together with a single store, achieving a 10× increase in scoring volume with only ~1 % latency growth.

Generative ad‑combination estimator : a two‑stage pruning strategy first scores candidates with a base model, selects the top N (N=3 online), then applies a context model based on a Transformer to capture product‑to‑product interactions. The architecture embeds store and product features, concatenates positional signals, and passes them through a Transformer to obtain context‑aware product embeddings, which are finally fed to a DNN for the combined score.

Cold‑start optimization : a Thompson‑sampling‑based Exploration & Exploitation (E&E) algorithm models each product’s pCTR as a Beta(a,b) distribution, adjusting the parameters (hyperP, hyperN) to give low‑exposure items higher exploration probability while keeping overall accuracy stable. This yields about 10 % random traffic without degrading precision.

These techniques have been deployed in multiple advertising scenarios. The heterogeneous mixing system improves Revenue‑Per‑Mille (RPM) by 4 %–15 % across scenes, and the cold‑start optimization adds 10 % random traffic without loss of accuracy. The paper concludes with a discussion of future directions, such as extending the shared‑representation and generative combination ideas to product‑creative pairing in broader advertising contexts.