The phenomenon where the statistical properties of a target variable change over time, degrading ML model performance.
Concept drift (also called dataset shift) refers to the change in the statistical relationship between input features and the target variable in a machine learning model over time. As the real world changes, the patterns a model learned during training may no longer hold — a fraud detection model trained on 2023 transaction patterns may fail to detect new fraud patterns in 2026. Concept drift is distinct from data drift (changes in input distribution) and semantic drift (changes in definition meaning).
Concept drift monitoring has become a standard practice in enterprise AI operations (MLOps). As AI models are deployed in production for months or years, their performance degrades as the world changes around them. In 2026, automated drift detection systems monitor model performance in real-time, triggering retraining when drift is detected. The semantic layer plays a role here too — when business definitions change (semantic drift), models trained on the old definitions experience a form of concept drift.
Concept drift is detected by monitoring model performance metrics (accuracy, F1, AUC) over time and comparing the distribution of model inputs and outputs against a reference window. Statistical tests like the Kolmogorov-Smirnov test, Page-Hinkley test, and ADWIN algorithm detect distribution shifts. When drift is detected, the model is retrained on recent data, fine-tuned, or replaced. Continuous learning systems update models incrementally as new data arrives.
A credit risk model trained on pre-pandemic data experiences severe concept drift when COVID-19 disrupts employment patterns. The model's AUC drops from 0.87 to 0.71 within three months. Automated drift monitoring detects the degradation, triggers an alert, and initiates retraining on recent data. The retrained model recovers to 0.84 AUC by incorporating the new economic patterns.