Intermediate · Evaluation

Model Drift

Visual diagram · (in preparation) · Math · (in preparation) · Worked example · 3 difficulty levels.

TL;DR. The degradation of a model's performance over time due to changes in the underlying data distribution.

Technical Definition

The degradation of a model's performance over time due to changes in the underlying data distribution.

How it works

Model drift is a critical issue in deployed machine learning systems, signifying that a model's predictive power diminishes as real-world data evolves away from its training data. This concept encompasses both 'covariate shift' (input data changes) and 'concept drift' (target variable relationship changes). Continuous monitoring and retraining are essential strategies to combat model drift and ensure sustained model accuracy in dynamic environments.