Case Study: Twitter achieves faster, higher-quality timeline ranking with TensorFlow

Twitter - Customer Case Study

Twitter, a global social media platform, struggled to surface the most relevant content for its users within their home timelines. Their original machine learning platform, built on Lua Torch, was inflexible, difficult to debug, and limited their ability to explore innovative model architectures to improve tweet ranking. To address these challenges, Twitter partnered with TensorFlow to rebuild their machine learning infrastructure.

By migrating to TensorFlow, Twitter was able to programmatically define complex, "split-network" model architectures, which improved prediction quality and reduced model training times by 22%. The TensorFlow platform also enabled warm-starting models, which allowed for faster iteration and led to a 0.4% reduction in relative log loss. This adoption of TensorFlow resulted in more relevant timelines for users, significant productivity gains for engineers, and a more stable, scalable production system for one of Twitter's largest machine learning applications.

Twitter

Yi Zhuang

Twitter