What is concept drift in classification systems?

Concept drift occurs when the statistical properties of input data or target labels change over time.

In production classification systems, models degrade because real-world data evolves. There are multiple types: covariate drift (input distribution changes), label drift (class proportions change), and concept drift (relationship between inputs and outputs changes). Detection methods include PSI (Population Stability Index), KL divergence, and online performance monitoring. Mitigation involves retraining pipelines, incremental learning, and adaptive models.