Home▸AI & Machine Learning▸Data Drift vs Concept Drift: How Production ML Models Quietly Stop Working

🧠 Data Drift vs Concept Drift

A 3D tunnel through time since model deployment: watch a frozen decision boundary drift out of sync with the true relationship in the data as data drift and concept drift accumulate — and trigger a retrain to see accuracy snap back.

AI & Machine Learning3DAdvanced60 FPS

A 3D tunnel through time since model deployment: a frozen decision boundary drifts out of sync with the true relationship in live data as data drift and concept drift accumulate, and a retrain snaps accuracy back.

🔬 What It Demonstrates

Each ring of glowing samples is a snapshot of production traffic at a point in time. The cyan plane is the model's frozen decision boundary; the twisting magenta ribbon is the true, live boundary. Samples turn red exactly where the two disagree — a real production misclassification.

🎮 How to Use

Raise data drift to shift the incoming feature distribution over time, raise concept drift to rotate the true relationship itself, and stretch the time horizon to see further into the future. Switch to "misclassified only" to isolate the errors, then hit retrain and watch the red points vanish.

💡 Did You Know?

Teams typically monitor for data drift with tests like the Population Stability Index or Kolmogorov–Smirnov on input features, and for concept drift by tracking live accuracy or calibration against delayed ground-truth labels — both trigger the same automated retraining pipeline this scene simulates.

⚙ Under the hood

machine learningdata driftconcept driftmodel retrainingproduction mlalgorithmsartificial intelligenceThree.js

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install