Meta learning of hyperparameters for parallel and distributed Gradient Boosted Decision Trees on big data

When Kinaxis trains its machine learning models, it does so on two time scales. Every week or so, it looks at around 10 billion sales records and tries to learn the rapidly-changing “parameters” that best describe this data. Every 3-6 months, it updates 10 thousand collections of “hyperparameters” that govern the parameter-learning process. Kinaxis would like to update the hyperparameters more frequently, but the update process is very expensive. In this project, we will find cheaper proxies for the hyperparameter update procedure that will allow much more frequent small updates, and thus better machine learning models. The basic idea is to look at historical relationships between hyperparameters, allowing us to calculate only a few updates and propagate the changes to all 10 thousand collections.

Faculty Supervisor:

Aaron Smith

Student:

Partner:

Kinaxis Inc.

Discipline:

Computer science

Sector:

Information and cultural industries

University:

University of Ottawa

Program: