Xiaohongshu's OneModel unifies ranking across 3 scenarios
Replaces 3 separate ranking systems with a single unified model that boosts user time (+0.33%) and ad CTR (+8.18%).
Researchers at Xiaohongshu (a social commerce platform with 200M+ users) have developed OneModel, a unified framework designed to replace fragmented ranking systems across multiple business streams like organic recommendations, advertising, and merchant services. By mapping heterogeneous user behaviors into shared event sequences and employing an action-oriented backbone, OneModel learns long-context user representations while using Scenario-aware Information Modulation to balance cross-stream knowledge transfer with stream-specific specialization.
In production deployment, the team implemented stratified user representation, multi-objective training, and optimized online serving with techniques like feature decomposition and shared user-tower computation. The system was deployed on Xiaohongshu’s Explore Feed, Feed Advertising, and Merchant Recommendation platforms, where it delivered measurable improvements: a +0.33% increase in time spent and +1.25% growth in engagement for the Explore Feed, a +3.43% lift in advertising value and +8.18% boost in CTR for Feed Advertising, and a +1.1867% increase in DGMV with a +2.1585% rise in GPM for Merchant Recommendation.
- OneModel unifies ranking across 3 disparate business streams (organic, ads, merchant) using shared event sequences and Scenario-aware Information Modulation
- Online A/B tests show +0.33% time spent, +1.25% engagement, +8.18% ad CTR, and +3.43% ad value lift
- Production deployment includes stratified user representation, multi-objective training, and graph-level inference optimization
Why It Matters
Proves unified multi-scenario ranking can reduce engineering overhead while improving core metrics across recommendation, advertising, and commerce.