Automated quality check and diagnosis for production model refresh

Abstract

As a data science project goes into the production stage, model maintenance to maintain model quality and predictive accuracy becomes a concern. Manual model maintenance by data scientists can become a time- and labor-intensive process, especially for large scale data science projects. An early warning system addresses this by performing systematic statistical and algorithmic checks for prediction accuracy, stability, and model assumption validity. A diagnostic report is generated that helps data scientists to assess the health of the model and identify sources of error as needed. Well-performing models can be automatically deployed without further human intervention while poor performing models trigger a warning or alert to the data scientists for further investigation and may be removed from production until the performance issues are addressed.