The gradual degradation of a machine learning model’s accuracy and predictive power over time as real-world data changes.
Model Drift (or Concept Drift) describes the phenomenon where the real-world data a model encounters in production begins to deviate from the historical data it was originally trained on.
As the environment changes (through shifting market trends, changing consumer behavior, or new regulations), the model's predictions become less accurate and less reliable.
Well-run MLOps pipelines monitor for model drift continuously. When drift crosses a defined threshold, the system automatically starts retraining, using fresh data from the pipelines built through data mobilization.
See also MLOps, Data Mobilization, Large Language Model, Fine-Tuning and Model Inference.