What is label drift and why does it matter?

medium

Answer

$Label drift = change in P(y) marginal distribution — class balance shifts over time.$
$Example: fraud rate rises from 1% to 3%.$
$P(y \mid x)$
$Detect by tracking observed class ratios in ground truth over time.$
$Handle via re-calibration (Platt / isotonic), threshold tuning, or class-weight re-adjustment.$
$(P(y \mid x))$

Check yourself — multiple choice

Same as concept drift
Change in P(y) marginal — class balance shifts over time; recalibrate + tune threshold; distinct from P(X) drift and $P(y \mid x)$ drift
Random
Not important

Label drift: P(y) shifts; recalibrate + retune threshold.

#monitoring#data-quality

215 interview questions in this topic.