文章背景与核心概要
双塔推荐模型作为 Netflix、Pinterest 和 Amazon 等各大平台的核心架构,在现代生产级推荐系统中得到了广泛应用。尽管它们因能够将用户和物品嵌入投影到低维空间而得到广泛部署,但其背后的底层理论属性在很大程度上仍未得到充分探索。本文作者对双塔推荐模型进行了正式的理论分析,确立了其渐近行为,证明了基于输入特征维度的更快收敛性,并提供了数值证据,表明该架构能够有效捕捉用户-物品属性交互,从而优于现有方法。
通过对两阶段推荐系统中应用的双塔模型进行研究,本文不仅建立了该模型的特定理论属性和统计保证,还通过合成数据和真实世界数据的实验,展示了双塔推荐器在封装物品与用户属性对评分影响方面的卓越性能。
迈向双Towers推荐模型的理论理解 (Towards a Theoretical Understanding of Two Tower Recommendation Models)
作者: Amit Kumar Jaiswal
标识符: arXiv:2403.00802 [cs.IR]
提交日期: 2024年2月23日(最近修订:2026年8月7日)
摘要 (Summary)
本文对双塔推荐模型(Two Tower recommendation models)进行了正式的理论分析。双塔模型是 Netflix、Pinterest 和 Amazon 等各大平台所采用的基石架构。虽然这些模型因能够将用户和物品嵌入(embeddings)投影到低维空间而得到广泛部署,但其潜在的理论属性在很大程度上仍未得到探索。作者确立了其渐近行为(asymptotic behaviors),证明了基于输入特征维度的更快收敛速度,并提供了数值证据,表明该架构能有效捕捉用户-Item 属性交互,从而优于现有方法。
This paper provides a formal theoretical analysis of Two Tower recommendation models, a cornerstone architecture used by major platforms like Netflix, Pinterest, and Amazon. While these models are widely deployed for their ability to project user and item embeddings into low-dimensional spaces, their underlying theoretical properties have remained largely unexplored. The author establishes asymptotic behaviors, demonstrates faster convergence based on input feature dimensionality, and provides numerical evidence that this architecture effectively captures user-item attribute interactions, outperforming existing methods.
抽象 (Abstract)
生产级的推荐系统严重依赖于在线媒体服务(包括 Netflix、Pinterest 和 Amazon)所使用的大规模语料库。这些系统通过利用双塔模型(两个深度神经网络)学习并投影在低维空间中的用户和物品嵌入来丰富推荐,从而促进其嵌入结构的构建以预测用户对物品的反馈。
尽管它在推荐领域广受欢迎,但其理论行为仍未得到全面的探索。我们研究了应用于两阶段推荐系统的双塔模型的渐近行为,该行为表现出向最优推荐系统的强收敛性。我们确立了双塔推荐系统的某些理论属性和统计保证。除了渐近行为外,我们还证明了采用双塔架构的推荐通过依赖输入特征的内在维度实现了更快的收敛。最后,我们通过数值实验表明,双塔推荐器能够封装物品和用户属性对评分的影响,与通过合成数据和真实世界数据实验进行的现有方法相比,取得了更好的性能。
Production-grade recommender systems rely heavily on a large-scale corpus used by online media services, including Netflix, Pinterest, and Amazon. These systems enrich recommendations by learning users' and items' embeddings projected in a low-dimensional space with two tower models (two deep neural networks), which facilitate their embedding constructs to predict users' feedback associated with items.
Despite its popularity for recommendations, its theoretical behaviors remain comprehensively unexplored. We study the asymptotic behaviors of the two tower model applied in two-stage recommenders that entail a strong convergence to the optimal recommender system. We establish certain theoretical properties and statistical assurance of the two tower recommender. In addition to asymptotic behaviors, we demonstrate that recommendation with two tower architecture attains faster convergence by relying on the intrinsic dimensions of the input features. Finally, we show numerically that the two tower recommender enables encapsulating the impacts of items' and users' attributes on ratings, resulting in better performance compared to existing methods conducted using synthetic and real-world data experiments.
元数据与资源 (Metadata & Resources)
| 类别 (Category) | 详情 (Details) |
|---|---|
| 学科 (Subjects) | 信息检索 (cs.IR);人工智能 (cs.AI) |
| 评论 (Comments) | 28页,3张图,11个表 |
| DOI | 10.48550/arXiv.2403.00802 |
| 许可 (License) | 查看许可 (view license) |
Category Details Subjects Information Retrieval (cs.IR); Artificial Intelligence (cs.AI) Comments 28 pages, 3 figures, 11 tables DOI 10.48550/arXiv.2403.00802 License view license
访问链接 (Access Links)
- 查看 PDF (View PDF)
- [HTML (实验性) (HTML (Experimental))]](https://arxiv.org/html/2403.00802v2)
- TeX 源码 (TeX Source)
- 音频摘要 (Audio Summary)
Access Links
提交历史 (Submission History)
- [v1] 2024年2月23日 星期五 21:11:55 UTC
- [v2] 2026年8月7日 星期五 17:00:54 UTC
- [v1] Fri, 23 Feb 2024 21:11:55 UTC
- [v2] Fri, 7 Aug 2026 17:00:54 UTC