A Service Oriented Architecture For Automated Machine Learning At Enterprise-Scale

Abstract

This thesis presents a solution architecture for productizing machine learning models in an enterprise context and, tracking the model’s performance to gain insights on how and when to retrain the model. There are two challenges which this thesis deals with. First, machine learning models need to be trained regularly to incorporate unseen data to maintain it’s performance. This gives rise to the need of machine learning model management. Second, there is an overhead in deploying machine learning models into production with respect to support and operations. There is scope to reduce the time to production for a machine learning model, thus offering cost-effective solutions. These two challenges are addressed through the introduction of three services under ScienceOps called ModelDeploy, ModelMonitor and DataMonitor. ModelDeploy brings down the time to production for a machine learning model. ModelMonitor and DataMonitor helps gain insights on how and when a model should be retrained. Finally, the time to production for the proposed architecture on two cloud platforms versus a rudimentary approach is evaluated and compared. The monitoring services give insight on the model performance and how the statistics of data change over time.