[9-4] Pixie: A system for Recommending 1+ Billion Items to 175+ Million Pinterest Users in Real-Time
题目：Pixie: A system for Recommending 1+ Billion Items to 175+ Million Pinterest Users in Real-Time
User experience in modern content discovery applications critically depends on personalized recommendations. However, a major challenge is how to build systems that can provide personalized recommendations from a massive pool of objects to a large number of users in real-time.
Here we present Pixie, a flexible graph-based real-time recommender system that we built and deployed at Pinterest. Today systems backed by Pixie contribute to more than half of all user engagement on Pinterest. Pixie harnesses the power of the Pinterest object graph, where users organize visual bookmarks, called pins, into collections, called boards. Given a set of query pins, Pixie uses fast random walks to traverse the bipartite graph and to identify pins most related to the query set. Pixie can make tens of thousands of recommendations per second, where each recommendation takes less than 60 milliseconds to complete.
We describe the algorithmic and system aspects of Pixie, examine several variants of random walk strategies, and discuss graph pruning strategies that lead to better quality of recommendations. Finally, we present experiments with offline data to demonstrate desirable properties of our system and explain how Pixie is used at Pinterest to help users discover and save ideas they love.
Dr. Zitao Liu is an applied scientist in Pinterest. Prior to joining Pinterest, he obtained his Ph.D. in Computer Science at University of Pittsburgh. During his Ph.D., his research work focuses on clinical data modeling, especially in time series domain. His research work was funded mostly by National Institutes of Health (NIH). He has more than 15 paper published in top ML/AI conferences, such as AAAI, CIKM, NAACL, ICDM, SDM, etc. He also worked on various challenging practical ML problems in Alibaba Group(2015), Yahoo! Labs(2014), eBay Research Lab(2013), and Google(2012).