A Hybrid Cross-Stage Coordination Pre-ranking Model for Online Recommendation Systems

TL;DR

This paper tackles the bottleneck in cascade-based recommender systems where lightweight pre-ranking models fail to align with downstream ranking, leading to sample selection bias and the Matthew effect. It introduces Hybrid Cross-Stage Coordination Pre-ranking (HCCP), combining Hybrid Multi-level Sample Construction and Hybrid Objective Optimization, including a novel Margin InfoNCE loss, to fuse information from upstream retrieval and downstream ranking. The approach explicitly leverages unexposed data across the entire stream and enforces both global/local consistency while enhancing long-tail item precision, achieving significant online gains (e.g., up to 14.9% UCVR and 1.3% UCTR on JD) with minimal latency impact. The work demonstrates strong offline and online performance improvements across Taobao and JD datasets, highlighting practical benefits for large-scale e-commerce recommendation systems and offering a framework transferable to other multi-stage architectures.

Abstract

Large-scale recommendation systems often adopt cascading architecture consisting of retrieval, pre-ranking, ranking, and re-ranking stages. With strict latency requirements, pre-ranking utilizes lightweight models to perform a preliminary selection from massive retrieved candidates. However, recent works focus solely on improving consistency with ranking, relying exclusively on downstream stages. Since downstream input is derived from the pre-ranking output, they will exacerbate the sample selection bias (SSB) issue and Matthew effect, leading to sub-optimal results. To address the limitation, we propose a novel Hybrid Cross-Stage Coordination Pre-ranking model (HCCP) to integrate information from upstream (retrieval) and downstream (ranking, re-ranking) stages. Specifically, cross-stage coordination refers to the pre-ranking's adaptability to the entire stream and the role of serving as a more effective bridge between upstream and downstream. HCCP consists of Hybrid Sample Construction and Hybrid Objective Optimization. Hybrid sample construction captures multi-level unexposed data from the entire stream and rearranges them to become the optimal guiding "ground truth" for pre-ranking learning. Hybrid objective optimization contains the joint optimization of consistency and long-tail precision through our proposed Margin InfoNCE loss. It is specifically designed to learn from such hybrid unexposed samples, improving the overall performance and mitigating the SSB issue. The appendix describes a proof of the efficacy of the proposed loss in selecting potential positives. Extensive offline and online experiments indicate that HCCP outperforms SOTA methods by improving cross-stage coordination. It contributes up to 14.9% UCVR and 1.3% UCTR in the JD E-commerce recommendation system. Concerning code privacy, we provide a pseudocode for reference.

Figures (8)

• Figure 1: Cascade architecture in recommendation system.
• Figure 2: Architecture of HCCP based on a three-tower model. (A) Hybrid sample construction module to obtain multi-level unexposed samples from upstream and downstream; (B) Consistency task to enhance the pre-ranking's rank ability on top items; (C) Hybrid optimization in cxr (ctr, cvr et.al) estimation tasks to improve the long-tail precision and mitigating SSB issue. (B) and (C) belong to hybrid objective learning, whose joint optimization achieves cross-stage coordination of pre-ranking.
• Figure 3: Ranking Sequence Construction. (A) Non-uniform sampling: randomly sampling items within uneven chunks. (B) Sequence arrangement for global and local consistency.
• Figure 4: The all-scenario click hitrate (ASH@K) of ranking and pre-ranking in the JD recommendation system.
• Figure 5: List-wise in-batch sampling. Top and tail items paired with other users can be sampled in different ratios.