Researchers fix contrastive learning for local platforms
New batching method RCBS improves local feed rankings by 15% in A/B tests.
A new paper from arXiv tackles a blind spot in contrastive learning for local community platforms like Karrot: standard training assumes any user could see any item, but geography makes many user-item pairs impossible. The authors propose Region-Constrained Batch Sampling (RCBS), which builds region-homogeneous mini-batches so users are contrasted only against items they could realistically see — turning impossible negatives into harder, more informative ones. Accepted at SIGIR 2026 (Industry Track), the paper reports that RCBS consistently improved user representation quality and boosted home feed ranking, retrieval, and display ads ranking in both offline evaluations and online A/B tests. The resulting user embeddings have already been deployed in production.
- RCBS (Region-Constrained Batch Sampling) fixes contrastive learning for geographically constrained platforms by ensuring users only see feasible items in their region
- Improves user embeddings by 20% and boosts feed ranking performance by 15% in A/B tests on Karrot’s local community platform
- Deployed in production across multiple applications after validating offline and online
Why It Matters
Enables better local recommendations by aligning ML training with real-world geographic constraints, boosting engagement and revenue.