MBGR: Multi-Business Prediction for Generative Recommendation at Meituan

Researchers from Meituan, China's leading food delivery and local services platform, have published a paper on MBGR (Multi-Business Generative Recommendation), a new AI framework designed to solve critical scaling problems in next-generation recommendation systems. Generative recommendation (GR) is an emerging paradigm that uses techniques like Semantic IDs (SIDs) and Next Token Prediction (NTP)—similar to how large language models work—to predict user preferences. However, applying this to a multi-business platform like Meituan, which offers food delivery, hotel bookings, and more, creates unique challenges. The team identified two major issues: the 'seesaw phenomenon,' where improving recommendations for one business type degrades performance for another, and 'representation confusion,' where a single semantic space fails to distinguish between items from different domains (e.g., a pizza vs. a hotel).

To address this, the MBGR framework introduces three key technical innovations. First, a Business-aware semantic ID (BID) module creates separate, domain-aware tokenization spaces to preserve the distinct semantics of items from different businesses. Second, a Multi-Business Prediction (MBP) structure provides dedicated prediction capabilities for each business line. Finally, a novel Label Dynamic Routing (LDR) module intelligently transforms sparse user interaction labels from across all services into dense, usable training signals. This architecture allows the model to learn complex, cross-business behavioral patterns—like a user ordering food after booking a hotel—without performance trade-offs.

The paper details extensive offline and online experiments conducted on Meituan's massive platform, confirming MBGR's effectiveness over previous methods. The system has already been deployed into production, marking a significant step in scaling generative AI techniques for real-world, multi-faceted e-commerce and service platforms. This work highlights the move beyond single-domain recommendation models toward unified, yet intelligently segmented, systems that can understand a user's holistic intent across a super-app's entire ecosystem.