Don't Contrast the Impossible: Region-Constrained Batching for Contrastive User Modeling on a Local Community Platform

2026-07-30

AI Summary

This work addresses a critical limitation in conventional contrastive learning within local community platforms, where user–item pairs that are geographically unobservable are erroneously treated as negative samples, thereby weakening the learning signal. To mitigate this issue, the authors propose Region-Constrained Batch Sampling (RCBS), a novel approach that constructs homogeneous mini-batches confined to specific geographic regions and retains only items that could realistically be exposed to users as negative samples. This strategy introduces harder and more informative contrastive signals by respecting real-world exposure constraints inherent to localized recommendation scenarios. RCBS is the first method explicitly designed for such exposure limitations, significantly enhancing user representation quality. It demonstrates consistent performance gains across multiple live deployment settings—including homepage recommendations, retrieval, and display ad ranking—and has been successfully integrated into production systems.

Abstract

Contrastive learning is widely used for user modeling in large-scale recommender systems, where standard in-batch negatives implicitly assume universal exposure that any user can be shown any item. On local community platforms such as Karrot, however, exposure is geographically constrained; many user-item pairs are impossible by design yet still treated as negatives during training, diluting the contrastive learning signal. We address this impossible negatives problem and propose Region-Constrained Batch Sampling (RCBS), a simple yet effective batching method that constructs region-homogeneous mini-batches so that users are contrasted primarily against items they could feasibly see. By replacing impossible negatives with feasible ones, RCBS naturally introduces harder and more informative negatives under realistic exposure constraints. With offline evaluations and online A/B tests, we show that RCBS consistently improves user representation quality and consequently enhances home feed ranking, retrieval, and display ads ranking. The resulting user embeddings have been deployed in production across various applications.

Problem

• contrastive learning
• impossible negatives
• geographic constraints
• user modeling
• local community platform

Innovation

• contrastive learning
• region-constrained batching
• impossible negatives
• user modeling
• local community platform