Presented By O’Reilly and Intel Nervana
Put AI to work
September 17-18, 2017: Training
September 18-20, 2017: Tutorials & Conference
San Francisco, CA

Escaping the forest, falling into the net: The winding path of Pinterest’s migration from GBDT to neural nets

Xiaofang Chen (Pinterest), Derek Cheng (Pinterest)
11:05am–11:45am Tuesday, September 19, 2017
Impact on business and society
Location: Imperial A Level: Intermediate
Secondary topics:  Case studies, Deep learning
Average rating: ****.
(4.67, 3 ratings)

Prerequisite Knowledge

  • A basic knowledge of machine learning systems

What you'll learn

  • Understand the challenges you may encounter when migrating from GBDT-based ML systems to neural nets and some solutions

Description

Pinterest is a catalog of the world’s ideas grounded in a content ecosystem of over 2B unique pins pinned to over 100B boards. To match users with this content, Pinterest has built a series of highly scalable recommender systems that have progressively allowed it to provide more and more personalized content. In 2017, the company completed its latest evolutionary step by migrating from a GBDT-based system to one using neural networks to predict various dimensions of user behavior.

Xiaofang Chen and Derek Cheng explore Pinterest’s recent transition to a TensorFlow-based system, covering the challenges and solutions to providing recommendations to over 160M monthly active users. While transitions like this can be frustrating for the teams undertaking them (and difficult for managers to justify), Xiaofang and Derek provide context to teams thinking about tackling a project like this so that they won’t be surprised by these gotchas and show how large gains are possible with a more flexible framework.

Topics include:

Benefits

  • User metrics gains
  • The flexibility to optimize to subpopulation performance (e.g., internationalization and other user demographics)
  • Increasing the overall robustness and intentionality of your pipeline
  • Extensibility of features and the ability to work with many more types of data

Challenges

  • Normalizing features
  • Feature engineering
  • Model design and tuning
  • Model debugging
  • Training and serving skew
  • Effective and efficient model serving
  • Model interpretability
Photo of Xiaofang Chen

Xiaofang Chen

Pinterest

Xiaofang Chen is a software engineer at Pinterest working on home feed ranking. Previously, Xiaofang was a software developer at Amazon. She holds a PhD in computer science from the University of Utah.

Photo of Derek Cheng

Derek Cheng

Pinterest

Derek Zhiyuan Cheng is software engineer on the discovery team at Pinterest, where he builds large-scale machine learning models and features to improve Pinterest’s personalization recommendation systems. Previously, he worked at Google Research, where he helped improve personalized search and recommendation systems for Google Play, News, and Google Plus. Derek has authored over 20 peer-reviewed articles published in prestigious conferences and journals for applied machine learning, information retrieval, and data mining. He holds a PhD with a focus on geosocial data mining from Texas A&M University.

Leave a Comment or Question

Help us make this conference the best it can be for you. Have questions you'd like this speaker to address? Suggestions for issues that deserve extra attention? Feedback that you'd like to share with the speaker and other attendees?

Join the conversation here (requires login)