Coordinating Predictions From Hundreds Of ML Models




Coordinating predictions from hundreds of ML models

21 August 2018

New York

Added 01-Jan-1970

Topic: Time Series Forecasting in Business

6:15pm - 6:45pm -Pizza & Drinks
6:45pm - 7:35pm - Talk
7:35pm - 8:15pm - Q&A and Networking

After completing a PhD in theoretical physics at King's College London, William worked at Penn State University as a research scientist focusing on computational and analytical models of physical phenomena, mathematical theories of gravitational wave propagation, and probabilities within quantum cosmological models. He continued as a research scientist at the Nijmegen University in the Netherlands where he received a Marie Curie Fellowship to continue his research on the theories of quantum cosmology. After his years of research, William joined Princeton Consultants as a systems integration specialist at a major transportation company, focusing on various networking and data platforms, notably around Positive Train Control. Since 2016, William has been a Data Scientist at Intent Media Inc., working to build and improve the core ML models that predict on-site user behavior. These models are used to make realtime decisions about user experience on most of the worlds largest travel sites.

At Intent Media we make real time decisions about how to treat users based on machine learning models. These models are trained daily on huge historical datasets, use real time data and return predictions in tens of milliseconds.

The real challenge is that we have hundreds of models, each of which have their own configurations, defaults, accuracy thresholds, decision strategy etc.

I will describe how we coordinate these models, how we automate the model publishing pipeline and how we monitor everything to ensure the best model is used every time.

I will keep the presentation at a reasonably high level, so anyone with experience (or interest) in product-ionizing machine learning models at scale should be able to take away something useful. No special background is required.