🧠 Data Drift vs Concept Drift

About this simulation

Each ring of points along the tunnel is a snapshot of production traffic at a given number of months since the model was deployed — nearest the camera is launch day, farthest away is today. A frozen cyan plane marks the decision boundary the model learned once, at launch, and never revisits. A twisting magenta ribbon marks the true separating boundary in the live data at each point in time. The two start aligned; watch them pull apart.

How it works

• Data drift slides the whole input distribution sideways over time — the incoming feature values move away from the region the model was trained on, without changing the true relationship between features and label.
• Concept drift rotates the magenta ribbon itself — the true relationship between features and label changes, so even points sitting exactly where they used to are now on the other side of "truth".
• A sample glows red when the frozen model's side of the cyan plane disagrees with the true label given by the magenta ribbon at that instant — i.e. it is a live production error.
• The accuracy, PSI (Population Stability Index) and boundary-shift readouts are recomputed live from every visible sample, exactly like a monitoring dashboard would.
• Retrain model now snaps the frozen cyan plane back onto today's true ribbon, instantly recovering accuracy — this is the automated retraining trigger described in the article.

Controls

• Data drift — how far the input feature distribution wanders from its launch-day position as months pass.
• Concept drift — how far the true feature→label relationship rotates away from what the model originally learned.
• Time horizon — how many months of production history the tunnel shows.
• Highlight — show every sample, or isolate only the ones the frozen model currently gets wrong.
• ↺ Retrain model now — re-fits the frozen boundary to the present day, resetting drift-driven error to zero. Drag to rotate, scroll to zoom.

Did you know?

Pure data drift with an unchanged relationship can sometimes be harmless — a correct rule still classifies shifted inputs correctly. It bites in practice because production models rarely extrapolate cleanly outside their training region. Concept drift is the more dangerous case: the ground truth itself has moved, so no amount of stable input data saves a stale model — only retraining on fresh labels does.