Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Request-Only Optimization for Recommendation Systems

Created by
  • Haebom

Author

Liang Guo, Wei Li, Lucy Liao, Huihui Cheng, Rui Zhang, Yu Shi, Yueming Wang, Yanzun Huang, Keke Zhai, Pengchao Wang, Timothy Shi, Xuan Cao, Shengzhi Wang, Renqin Cai, Zhaojie Gong, Omkar Vichare, Rui Jian, Leon Gao, Shiyan Deng, Xingyu Liu, Xiong Zhang, Fu Li, Wenlei Xie, Bin Wen, Rui Li, Lu Fang, Xing Liu, Jiaqi Zhai

Outline

This paper presents Request-Only Optimizations (ROO), a novel paradigm for efficient training and storage of deep learning recommendation models (DLRM), one of the world's largest machine learning applications. ROO processes user requests as data units, saving data storage space compared to traditional user impression-based processing. It also eliminates redundant computation and communication across multiple impressions within a request, enabling extended neural network architectures (e.g., generative recommendation models) to better capture user interest signals. To achieve this, we design an integrated approach that integrates data (request-only data), infrastructure (request-only data processing pipelines), and model architecture (request-only neural network architectures).

Takeaways, Limitations

Takeaways:
Reduce data storage space: Eliminate redundant data by processing user requests on a per-data basis.
Improved learning efficiency: Eliminate redundant computation and communication across multiple exposures within a request.
Improved model performance: More accurately captures user interest signals through an extended neural network architecture.
Expanding the applicability of new architectures such as generative recommendation models.
Limitations:
Lack of specific details on the application and performance evaluation of the ROO paradigm in real industrial environments.
Further research is needed on the generalizability of ROO to different types of recommender systems.
Consider the challenges and costs of building and managing a request-only data processing pipeline.
👍