Pinterest Home Feed 统一轻量级评分。双塔的方法

Dafang He | Software Engineer, Home Candidate Generation; Andrew Liu; Dhruvil Deven Badani | Software Engineer, Homefeed Ranking; Poorvi Bhargava; Sangmin Shin |Engineering Manager, Home Ranking; Duo Zhang | Engineering Manager, Candidate Generation; and Jay Adams | Software Engineer, Inspire

Intro

介绍

Pinterest is a place where users (Pinners) can save and discover content from both web and mobile platforms, and where increasingly Creators can publish native content right to Pinterest. We hold billions of content (Pins) in our corpus and serve personalized recommendations that inspire Pinners to create a life they love. One of the key and most complicated surfaces for Pinterest is the home feed, where Pinners will see personalized feeds based on their engagement and interests. In this blog, we will discuss how we unify our light-weight scoring layer across the various candidate generators that power home feed recommendations.

Pinterest是一个用户（Pinners）可以保存和发现来自网络和移动平台的内容的地方，而且越来越多的创作者可以在Pinterest上发布本地内容。我们的语料库中拥有数十亿内容（Pins），并提供个性化的推荐，激励品客创造他们喜爱的生活。Pinterest的关键和最复杂的表面之一是主页，品客将在这里看到基于他们的参与和兴趣的个性化的内容。在这篇博客中，我们将讨论如何在为主页推荐提供动力的各种候选生成器中统一我们的轻量级评分层。

Motivations

动机

Home feed is what Pinners see first when they open the Pinterest app. To give relevant and diverse recommendations, we use a recommendation system comprising many different sources. One major source, for example, is Pixie[1], which is based on a random walk of the bipartite pin-board graph. Based on the Pixie platform, we are able to generate multiple different sources, some directly returning pins from a random walk based on the engagement history, and some based on the pins retrieving from boards that returned from Pixie random walk. In addition to Pixie, we also have recommendation sources that take in the topics or based on embeddings. These candidate generators usually have their own light-weight scoring model, w...