What Is Model Monitoring and How Does It Work?

It involves constantly watching the inputs, outputs, and performance of a deployed model to make sure it stays working. Issues such as data drift and accuracy decay, which can quietly degrade a model after launch, are caught through monitoring.

Data drift occurs when the input data a model is fed in a live environment differs from the data used to train it. The model then faces inputs it was not prepared for, which can quietly degrade its accuracy even though the model itself has not changed.

Data drift is a shift in the distribution of the inputs. Concept drift is a change in the input-output relationship. Both hurt accuracy but are detected and dealt with in different ways, which is why monitoring tracks both.

Because the environment it operates in keeps changing. User behavior, data, and conditions shift over time, so a model that was accurate at launch can drift out of step with reality and make poor predictions even though its code and weights are unchanged.

Evaluation tests a model against known cases, usually before deployment. Monitoring watches the model continuously after deployment, where true answers often arrive late, so it relies on drift and operational signals as well as accuracy to catch problems.