Domain adaptation

Summary

Training on one distribution and deploying on a related but shifted one is the normal case, not the exception: a fraud model meets new fraud, a medical model meets a new hospital’s scanner, a speech model meets a new accent. Domain adaptation transfers a predictor from a source distribution $P_{s} (X, Y)$ to a target $P_{t} (X, Y)$ where the label space is usually the same but the inputs, prevalences, or input-label relationship have moved. The first job is to name which of those moved, because that decides whether the problem is even solvable from the data you have.

Shift types

Covariate shift: $P (X)$ changes while $P (Y ∣ X)$ is stable.
Label shift: $P (Y)$ changes while $P (X ∣ Y)$ is stable.
Concept shift: $P (Y ∣ X)$ changes; unlabeled target data alone is generally not enough.

Naming the assumed shift determines which correction is defensible.

Learning objective: distinguish covariate, label, and concept shift by the factor that changes, then decide whether unlabeled target inputs contain enough evidence to adapt.

Assumption map

What changed, what stayed fixed, and what can target inputs reveal?

Read it this way: read each row from left to right. A correction is justified only after naming both the factor that changed and the one assumed fixed. Target inputs can expose a changed input mix and, with identifiability conditions, a changed class mix; they cannot reveal a new label rule by themselves. Original synthesis checked against Sugiyama et al. (2007), Lipton et al. (2018), and Ben-David et al. (2010).

Approaches

Importance weighting under a covariate- or label-shift assumption
Fine-tuning on a small labeled target set
Feature alignment with a discrepancy or adversarial objective
Self-training with confidence and calibration controls
Domain-specific normalization or adapters
Robust optimization across the environments you can observe

Evaluation

Use a true target-domain holdout and report the slices that matter. Validate calibration, not just ranking or accuracy, and measure negative transfer: adaptation can help the aggregate while hurting a target subgroup or eroding source performance.

In an interview

Define source, target, labels, and how much target supervision you have.
State the shift assumption.
Establish source-only and target-labeled baselines.
Pick the simplest method the evidence justifies.
Monitor drift and collect the labels that distinguish concept shift from covariate shift.

Common confusions

“Align the feature distributions and the task transfers.” Alignment can mix classes or erase the predictive structure you needed.
“Unlabeled target data solves domain shift.” Not when the label relationship itself changed.
“Fine-tuning always helps.” A small, biased target set can cause negative transfer and calibration failure.