Concept Drift

ELI5 - The Vibe Check

Concept drift is a sneakier cousin of data drift: the inputs might look exactly the same, but the actual relationship between those inputs and the right answer has changed underneath the model's feet. A spam filter trained on old spam patterns can fail not because emails look different, but because spammers changed their tactics, so what used to reliably mean spam doesn't mean that anymore.

Real Talk

Concept drift occurs when the statistical relationship between a model's input features and the target variable changes over time, meaning the function the model learned no longer reflects reality, even if the input distribution itself looks stable. It's distinct from data drift, which is a shift in the inputs themselves, and is typically caught by monitoring live prediction accuracy against ground truth as it becomes available, not just by watching input distributions.

When You'll Hear This

"Accuracy is dropping but the input data looks totally normal, that smells like concept drift, the relationship itself changed." / "Fraud patterns evolve constantly, so we treat concept drift as a given and retrain the model every week."