下一层次的个性化：如何通过16k+终身用户行为提升Pinterest的推荐
[
[
Xue Xia | Machine Learning Engineer, Home Feed Ranking; Saurabh Vishwas Joshi | Principal Engineer, ML Platform; Kousik Rajesh | Machine Learning Engineer, Applied Science; Kangnan Li | Machine Learning Engineer, Core ML Infrastructure; Yangyi Lu | Machine Learning Engineer, Home Feed Ranking; Nikil Pancha | (formerly) Machine Learning Engineer, Applied Science; Dhruvil Deven Badani | Engineering Manager, Home Feed Ranking; Jiajing Xu | Engineering Manager, Applied Science; Pong Eksombatchai | Principal Machine Learning Engineer, Applied Science
Xue Xia | 机器学习工程师，主页动态排名；Saurabh Vishwas Joshi | 首席工程师，ML平台；Kousik Rajesh | 机器学习工程师，应用科学；Kangnan Li | 机器学习工程师，核心ML基础设施；Yangyi Lu | 机器学习工程师，主页动态排名；Nikil Pancha | （前）机器学习工程师，应用科学；Dhruvil Deven Badani | 工程经理，主页动态排名；Jiajing Xu | 工程经理，应用科学；Pong Eksombatchai | 首席机器学习工程师，应用科学
Background
背景
The Pinterest home feed is crucial for Pinner engagement and discovery. Pins on the home feed are personalized using a two-stage process: initial retrieval of candidate pins based on user interests, followed by ranking with the home feed (Pinnability) model. This model — a neural network consuming various pin, context, and user signals — predicts personalized pin relevance to improve user experience. Its architecture is illustrated in the following figure.
Pinterest的首页动态对Pinner的参与和发现至关重要。首页动态上的Pins通过两阶段过程进行个性化：基于用户兴趣的候选Pins的初步检索，随后使用首页动态（Pinnability）模型进行排名。该模型——一个消耗各种Pin、上下文和用户信号的神经网络——预测个性化Pin的相关性，以改善用户体验。其架构在下图中进行了说明。
In 2023, we published TransAct (arXiv), where we use transformers to model the Pinner’s last 100 actions in Pinnability. One shortcoming of that approach is that we are unable to model the user’s lifelong behavior on Pinterest.
在2023年，我们发布了TransAct（arXiv），我们使用变换器来建模Pinner在Pinnability中的最后100个动作。这种方法的一个缺点是我们无法建模用户在Pinterest上的终身行为。
T...