Delivery and mobility / Ride-hailing and delivery
Grab: A shared user model across several recommendation services
- Company
- Grab
- Country
- Not classified
- Adoption stage
- In operation
- Source published
- Date basis
- The date the source was published. It can differ from the date adoption started.
- How the source was checked
- Read the full source text
The work problem
Grab's recommendation models relied on user features engineered separately for each task. Teams had to process food orders, rides and spending independently, and struggled to capture the sequence of user actions in the app. A reusable user representation across services was needed.
Technology and data
Grab built a Transformer-based user model that jointly learns from tabular transaction and user-attribute data and time-ordered app views and clicks. Modality-specific adapters and sequence handling yield embeddings for long-term preferences and recent intent. A Ray-based pipeline separates CPU data preparation from GPU model execution, refreshes embeddings for millions of active users daily and supplies them to downstream models.
Results
Grab said ad optimization, dual app prediction, fraud detection and churn probability already benefit from the model. Teams can add common embeddings to existing models or reuse them in new model development. The article describes reduced development time and improved predictions, but gives no service-specific online experiment figures or monetary savings.
Limits and open questions
Task-specific accuracy, online experiment effects and cost savings were not established. Unifying data streams across all services and expanding to more entity types remain future improvements.
Sources
- Grab | 여러 추천 서비스의 공통 사용자 모델engineering.grab.com, Accessed
Compiled from public sources. These are not results from ATF Works customers.