Model Degradation
Model Degradation is the decline in a machine learning model's performance, accuracy, reliability, or effectiveness over time or when operating under changing data and real-world conditions.
What is Model Degradation?
Model degradation can occur when the data used in production differs from the data used during training, patterns in the environment change, or the model's underlying assumptions no longer hold. It may result from factors such as data drift, concept drift, changing user behavior, or outdated training data.
Why is Model Degradation Important?
A degraded model can produce increasingly inaccurate or unreliable predictions without obvious changes to the model itself. Monitoring model performance and detecting degradation helps teams identify when a model needs investigation, recalibration, retraining, or replacement.
Common use cases
Model degradation is commonly monitored in fraud detection, recommendation systems, forecasting, predictive maintenance, credit scoring, and other production machine learning applications.