解锁高效广告检索：Pinterest Ads中的离线近似最近邻
[
[
Authors (non-ordered): Qishan(Shanna) Zhu, Chen Hu
Acknowledgements: Longyu Zhao, Jacob Gao, Quannan Li, Dinesh Govindaraj
作者（无序）：Qishan(Shanna) Zhu, Chen Hu
致谢：Longyu Zhao, Jacob Gao, Quannan Li, Dinesh Govindaraj
Introduction
介绍
In the evolving landscape of advertising, the demand for real-time personalization and dynamic ad delivery has made Online Approximate Nearest Neighbors (ANN) a mainstream method for ad retrieval. Pinterest primarily employs online ANN to swiftly adapt to users’ behavior changes (depending on their age, location and privacy settings), thereby enhancing ad responsiveness and relevance.
在不断发展的广告环境中，对实时个性化和动态广告投放的需求使得在线近似最近邻（ANN）成为广告检索的主流方法。Pinterest主要采用在线ANN迅速适应用户的行为变化（取决于他们的年龄、位置和隐私设置），从而增强广告的响应性和相关性。
However, offline ANN is also a valuable option, particularly when large-scale data processing, efficient resource utilization, and cost-effective operations are critical. By precomputing candidates offline, this approach is ideal for scenarios that require high throughput and low-latency query responses and relatively static query context. This article explores suitable use cases for offline ANN, outlines its advantages and disadvantages, shares insights from our experiences, and illustrates its application within Pinterest. We will also discuss potential future enhancements.
然而，离线ANN也是一个有价值的选择，特别是在大规模数据处理、高效资源利用和成本效益操作至关重要时。通过离线预计算候选项，这种方法非常适合需要高吞吐量和低延迟查询响应以及相对静态查询上下文的场景。本文探讨了离线ANN的适用用例，概述了其优缺点，分享了我们的经验见解，并说明了其在Pinterest中的应用。我们还将讨论潜在的未来增强。
Problem Statement
问题陈述
Pinterest has successfully applied Online ANN to fetch from its large ads inventory, which has brought double digits gains on ads quality metrics across surfaces. However, we are encountering challenges as the ads inventory continuously expands. It is imperative to maintain a neutral imp...