Machine learning has allowed Twitter to drive engagement, promote healthier conversations, and deliver catered advertisements. Cibele Montez Halasz and Satanjeev Banerjee describe one of those use cases: timeline ranking. They share some of the optimizations that the team has made—from modeling to infrastructure—in order to have models that are both expressive and efficient. You’ll explore the feature pipeline, modeling decisions, platform improvements, hyperparameter tuning, and architecture (alongside discretization and isotonic calibration) as well as some of the challenges Twitter faced by working with heavily text-based (sparse) data and some of the improvements the team made in its TensorFlow-based platform to deal with these use cases. Join in to gain a holistic view of one of Twitter’s most prominent machine learning use cases.
Cibele Montez Halasz is a machine learning engineer at Twitter Cortex, where she helps to build Twitter’s deep learning platform. Previously, Cibele was a data scientist and systems design engineer at Apple and a product applications engineer at Analog Devices, where she built machine learning algorithms that use smartphone sensors to understand a person’s behavior. Cibele holds an MS in electrical engineering with an emphasis on computer vision and machine learning from the California Institute of Technology and a BS in electrical engineering and physics from Stanford University.
Satanjeev ‘Bano’ Banerjee is a machine learning engineer at Twitter working on Tweet relevance ranking. Previously he worked as a data scientist on Twitter’s Advanced Analytics team, and as a backend engineer within the consumer engineering team. He holds a PhD from CMU; his dissertation was on active learning in the context of meeting understanding and summarization. Bano’s interests outside of work include volunteering his data analysis skills at local non-profits, and training towards a marathon.
©2019, O'Reilly Media, Inc. • (800) 889-8969 or (707) 827-7019 • Monday-Friday 7:30am-5pm PT • All trademarks and registered trademarks appearing on oreilly.com are the property of their respective owners. • email@example.com