Global User Modeling and Explicit Interest Transfer Framework for Meituan Home Page Recommendation

This article details the development and evolution of global user modeling techniques in Meituan's home page recommendation system. The recommendation team adopted a multi‑stage, progressive exploration to incorporate multi‑venue and multi‑channel user interaction data into both recall and ranking modules, addressing severe negative transfer caused by strong cross‑venue, cross‑business, and spatio‑temporal correlations.

1. Background

Meituan's home page recommendation serves tens of millions of users daily across diverse services such as food delivery, travel, e‑commerce, and more. Different venues exhibit distinct user behavior patterns and business characteristics, leading to data sparsity and bias in the original recommendation pipeline that relied solely on home‑page interactions.

Global user modeling aims to transfer user interests from source domains (e.g., search, “golden area” venues) to the target home‑page domain, thereby enriching the signal space and mitigating sparsity.

2. Exploration and Deployment of Global User Modeling

Stage 1 – Optimizing Global Recall Strategy : Expanded recall data sources from home‑page only to multi‑venue and multi‑channel behaviors (click, search, payment), quickly validating feasibility.

Stage 2 – Introducing Global Signals into Model Training : Re‑labeled training samples by injecting payment data from source venues as positive signals, while applying business‑level hierarchical weights to alleviate negative transfer.

Stage 3 – Explicit Interest Transfer Framework (EXIT) : Proposed a novel explicit cross‑domain recommendation framework that uses interest combination labels and a Scene Selector Network (SSN) to selectively transfer useful source‑domain interests based on fine‑grained spatio‑temporal contexts.

Stage 4 – Unified Global Modeling and Perception Enhancement : Built a unified sample pipeline covering full‑chain, full‑supply, and full‑domain data, added both positive and negative samples from external venues, and constructed comprehensive global features.

Stage 5 – Generative Recommendation Paradigm (Future) : Plans to combine large language models with global behavior sequences for a generative recommendation approach.

3. EXIT Framework Details

The EXIT architecture consists of a multi‑task interest modeling network, interest combination labels (ICL), and a Scene Selector Network (SSN). ICL provides explicit supervision for cross‑domain interest transfer, while SSN predicts transfer probabilities conditioned on user, item, and contextual embeddings.

Mathematically, the final transferred interest is expressed as:

Interest_target = Interest_target_tower + TransferProb * Interest_source_tower

SSN processes concatenated embeddings of user, item, and scene features to output the transfer probability.

4. Offline and Online Evaluation

Extensive experiments compared EXIT against classic and cross‑domain baselines. EXIT achieved higher offline AUC, improved online CTCVR, and reduced negative feedback rate (NFR). Ablation studies confirmed the importance of each module, especially the ICL.

Visualization of transfer probabilities for the food‑delivery business showed three daily peaks corresponding to breakfast, lunch, and dinner, demonstrating effective scene‑aware transfer.

5. Summary and Outlook

The multi‑stage global user modeling has substantially boosted Meituan's home‑page recommendation performance, alleviating data sparsity and negative transfer. The EXIT framework has been deployed at scale and accepted as a paper at CIKM 2024 ("EXIT: An EXplicit Interest Transfer Framework for Cross‑Domain Recommendation"). Future work includes incorporating external click signals, extending EXIT to non‑overlapping item domains, and exploring generative recommendation models.