Contextualized Recommendation in Meituan Takeaway: Segmented & Unified Modeling, Long‑Sequence Retrieval, and Multi‑Expert Networks

Meituan Takeaway’s recommendation team presents a contextual modeling framework for the food‑delivery scenario. The approach, called “Segmented and Unified Model”, first partitions user contexts (time, location, entry page, business type) and then shares knowledge across similar contexts through a unified model.

The paper describes three core components: (1) user behavior sequence modeling that retrieves the most relevant historical actions for the current context, (2) a multi‑expert network that learns specialized representations for each context, and (3) engineering optimizations for long‑sequence handling.

1. Introduction – In takeaway services, a user is not only a person but also a collection of context‑dependent demands. Different times, places, and business types lead to distinct user needs, motivating contextual recommendation.

2. Problems and Challenges – Geographic and cultural constraints cause large variations in order patterns between weekdays and weekends. Traditional models that treat all behavior as a single sequence struggle to capture these nuances.

3. Contextual Intelligent Traffic Distribution

3.1 Contextual Long‑Sequence Retrieval – The system builds on deep learning models such as DIN, DIEN, and MIMN, but adds a retrieval step that treats the target item as a query and extracts a context‑relevant subsequence from the user’s long history. Fine‑grained micro‑behaviors (≈70 types, aggregated into 12 intents) are used to enrich the sequence.

3.2 Contextual Multi‑Expert Network – A Mixture‑of‑Experts (MoE) architecture is employed, with dense and sparse MMOE variants. Experts are activated based on strong contextual features (city, time, entry). An AutoAdapt mechanism shares a universal expert while allowing domain‑specific experts to be dynamically gated.

3.3 Engineering Optimizations – Data transmission is reduced by de‑duplicating tag‑id features, and model inference is accelerated by moving embeddings to GPU hash tables and folding identical user‑side sub‑graphs.

4. Experimental Results – Offline experiments show gains of +0.30 pp CTR GAUC and +0.52 pp CXR GAUC from the segmented‑unified model. Online metrics improve UV_RPM (+0.70 % to +1.77 %), UV_CXR (+0.87 %), PV_CTR (+0.70 % to +0.89 %), exposure novelty (+1.51 %), and first‑order purchase share (+1.29 %). Dense MMOE with 4 experts and sparse Top‑K MMOE further boost performance while keeping latency within constraints.

5. Summary and Outlook – The Cube concept enables continuous exploration of new contexts and cold‑start mitigation by leveraging similar context user pools. Future work includes unified multi‑attribute retrieval, parameter‑efficient sparse expert designs, and AutoML‑driven expert selection.