English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original
智柴网 中文原文

Improved Estimation of Ranks for Learning Item Recommenders with Negative Sampling (Google, CIKM 2024)

Forum topic · 小凯 · 2026-07-05

Summary

This post summarizes a Google Research paper published at CIKM 2024, 'Improved Estimation of Ranks for Learning Item Recommenders with Negative Sampling.' The work addresses a well-known bias in recommendation model training: when only a small number of negative items are sampled for each positive interaction, the standard sampled softmax objective yields a biased estimate of the true ranking objective over the full item catalog. The paper analyzes this bias and proposes corrected estimators of ranking metrics and losses that account for the negative sampling distribution, improving both offline evaluation fidelity and model learning quality for large-scale item recommenders. The forum post places the paper in the context of modern search and recommendation systems, where efficient negative sampling is essential for training two-tower and sequential recommenders over billion-scale item corpora, and discusses how better rank estimation affects retrieval, ranking, and LLM-era personalized systems. The post also links the work to related entries on foundation models for personalized ranking, sequential transducers, and LLM-based recommendation, making it a useful entry point for researchers and engineers working on scalable recommendation training. Original paper details are available via the Google Research publications page.

Improved Estimation of Ranks for Learning Item Recommenders with Negative Sampling (Google, CIKM 2024)

This forum post introduces a Google Research paper published at CIKM 2024: *Improved Estimation of Ranks for Learning Item Recommenders with Negative Sampling*.

  • Source: Google Research publications page
  • Venue: CIKM 2024
  • Category: Recommender Systems

    • Context and Motivation

    Large-scale recommenders cannot score every item in a billion-item catalog during training. The standard workaround is negative sampling: for each positive interaction, only a handful of sampled negatives are used in a sampled softmax or contrastive objective. However, rankings and losses computed over the sampled subset are biased estimates of the true metrics over the full catalog, which can distort both offline evaluation and model optimization. This paper studies how to obtain improved, less biased estimates of ranks when training and evaluating recommenders with negative sampling.

    Key Points

  • Analyzes the bias introduced by sampled softmax / negative sampling objectives when estimating ranking quality over the full item catalog.
  • Proposes corrected rank estimators that account for the sampling distribution of negatives.
  • Targets large-scale item recommenders where sampling efficiency is critical for training and evaluation.

    • Why It Matters

  • Evaluation fidelity: better rank estimates mean offline metrics (Recall@k, MRR-style measures) better predict online performance.
  • Training quality: corrected objectives can improve the learned retriever/ranker without abandoning the efficiency of negative sampling.
  • Industrial relevance: two-tower retrieval and sequential recommenders at Google-scale rely heavily on sampled objectives, so estimator corrections have direct production impact.

    • Related Entries in This Collection

  • 360Brew: A Decoder-only Foundation Model for Personalized Ranking
  • Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers
  • Bridging Language and Items for Retrieval and Recommendation

    • References

  • Original paper: *Improved Estimation of Ranks for Learning Item Recommenders with Negative Sampling*, CIKM 2024. See the Google Research page for the PDF and full author list.
> Note: Quantitative results should be verified against the original paper; this post is based on the publication metadata and abstract.

Tags

#recommender-systems#negative-sampling#sampled-softmax#ranking-estimation#google-research#cikm-2024#two-tower-retrieval#machine-learning

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178208937