Summary
In this chapter, you have first been introduced to the underlying foundations of model drift. You have seen that model drift and a decrease in model performance over time are to be expected in ML models in a real-life environment.
Decreasing performance can generally be attributed to drifting data, drifting concepts, or model-induced problems. Drifting data occurs when data measurements change over time, but the underlying theoretical concept behind the model stays the same. Concept drift captures problems of those theoretical underlying foundations of the learned processes.
Model- and model retraining-related problems are generally not considered standard reasons for drift, but they should still be monitored and taken seriously. Depending on your business case, relearning—especially if monitoring is lacking—can introduce large problems with ML systems.
Data drift can generally be measured well by using descriptive statistics. Concept drift is often harder...