LongRetriever: Towards Ultra-Long Sequence based Candidate Retrieval for Recommendation

2025-08-21

AI Summary

Existing industrial recommendation systems struggle to effectively model ultra-long user behavior sequences during the recall stage, leading to insufficient fine-grained interest representation. To address this, we propose LongRetriever—the first framework to incorporate ultra-long sequence modeling into recall. It introduces a candidate-aware contextual training paradigm to enable fine-grained interaction between behavioral sequences and candidate items; designs a multi-context retrieval mechanism to ensure consistency between training and online serving; and integrates contextual learning, candidate-specific interaction, and search-based multi-context retrieval within a scalable online optimization architecture. Deployed on a large-scale e-commerce platform, LongRetriever significantly improves online click-through rate (CTR) and conversion rate (CVR), and has been stably serving billions of users in production.

Technology Category

• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval
• Search and Optimization: Learning to Search
• Machine Learning: Learning Preferences or Rankings

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: Economics and fairness of platforms and recommendation systems

Abstract

Precisely modeling user ultra-long sequences is critical for industrial recommender systems. Current approaches predominantly focus on leveraging ultra-long sequences in the ranking stage, whereas research for the candidate retrieval stage remains under-explored. This paper presents LongRetriever, a practical framework for incorporating ultra-long sequences into the retrieval stage of recommenders. Specifically, we propose in-context training and multi-context retrieval, which enable candidate-specific interaction between user sequence and candidate item, and ensure training-serving consistency under the search-based paradigm. Extensive online A/B testing conducted on a large-scale e-commerce platform demonstrates statistically significant improvements, confirming the framework's effectiveness. Currently, LongRetriever has been fully deployed in the platform, impacting billions of users.

Problem

• Modeling ultra-long user sequences for recommendation systems
• Incorporating ultra-long sequences into candidate retrieval stage
• Enabling candidate-specific interaction with user sequences

Innovation

• In-context training for candidate-specific interaction
• Multi-context retrieval ensuring training-serving consistency
• Ultra-long sequence incorporation in retrieval stage