Machine learning model maintenance after shift detection

In machine learning, we talk about model degradation when dataset shifts and prediction shifts occur. When these shifts are detected, it is the moment to adjust the model. We refer to model maintenance as the strategies we can adopt to adjust the model after a shift occurs so it can be deployed again and continue providing predictions.

In this nugget, we will study supervised and unsupervised methods to adjust a model after dataset shifts and prediction shifts occur. We will depart from a roadmap of model maintenance highlighting the stages of shifts, monitoring, and adjustment. The strategies for the model update include single corrections, recalculation of the existing models, and transfer learning approaches. We will gather the insights of these different methods and the concepts of dataset shifts to discuss when the model can be successfully adjusted. This nugget will guide the learner through these concepts with a simulated industrial process involving sensors to predict a quality compound of interest.