Using ML to Power Live-stream Recommendations at a Global Scale

Abstract

Twitch is a premier live social video platform with millions of unique streamers and tens of millions of daily viewers (who watched 1.3 trillion minutes of content in 2021). Recommendations are a major means of connecting viewers to streamers at Twitch. In this talk, I will discuss the challenges of recommending live video content, and how it differs from traditional Video on Demand recommendations. I will  then present our neural network based algorithms for generating live stream recommendations, highlighting how these algorithms tackle some of the challenges associated with live content. I’ll present methods, offline experimentation, and evaluation metrics for live-stream recommendation scenarios. Finally, I’ll talk about the solution’s success in a sitewide A/B test that increased recommended minutes watched by 7.9%. At the start of the talk, I’ll also discuss my journey into ML based product development, and present key considerations for applied-ML research that powers products operating at a global scale.

For more info, please follow this link.