On Day 1, we looked at Data Quality and Model Quality drift. Today we cover the remaining two monitors from Model Monitor: Bias Drift and Feature Attribution Drift. These two are handled by SageMaker Clarify. Clarify is a service that calculates model bias before training, post-training model bias, and prediction explainability. Today's topic extends that capability to continuous monitoring of deployed endpoints.
Why do we need to monitor bias and explainability during operations? A model that was fair at training time can become unfair to certain groups as incoming data distribution changes