🌊  Label Drift (a.k.a. Target Drift)

Label Drift (a.k.a. Target Drift)#

A change over time in the distribution of the target variable.

Important

✨ AI-generated content. This page was written with the assistance of an AI language model and is provided as a learning aid. Despite careful review, it may still contain mistakes, omissions, or out-of-date information. Whether you are new to the topic, a team lead, or a senior practitioner, treat it as a starting point rather than an authoritative reference: read it critically and independently verify anything you act on (code, commands, figures, and factual claims) against official documentation and primary sources before relying on it.

What it is#

Label drift (target drift) is a change in the distribution of the target itself — the class balance or outcome mix shifts between training and production, \(p_{\text{train}}(y) \neq p_{\text{prod}}(y)\), even when the feature-to-label relationship may be unchanged. A fraud rate that creeps from 1% to 3% is label drift.

How it differs#

Like covariate drift it is a dataset shift, but it moves p(y) rather than p(x) or p(y | x). Because most classifiers implicitly assume the base rate they trained on, a shifted target distribution can throw off calibrated probabilities and thresholds even if each input still maps to the right answer.

Detecting and fixing it#

Monitor the label or prediction distribution over time (PSI on predicted classes, tracked class proportions). Fixes include recalibrating decision thresholds to the new base rate, reweighting or resampling to the current mix, and retraining on recent labels.


Theme: Distribution Shift & Drift  ·  All terminology



See also

Source article Adapted (context, re-expressed) in our own words from: Label Drift (a.k.a. Target Drift) (insightful-data-lab.com).

Tags: purpose: reference topic: terminology level: advanced