Does adaptation always require target labels?

No. Some methods use unlabeled target examples, although evaluation still benefits from trustworthy outcomes.

Is fine-tuning the same thing?

Fine-tuning can be an adaptation technique, but it is not the only one.

Can adaptation make a model worse?

Yes. Negative transfer occurs when reused information harms the target task.

AI Glossary / AI and machine learning / Intermediate

Domain adaptation

Definition

Domain adaptation adjusts a machine-learning system so that knowledge learned from one data setting can be useful in another. It addresses differences between the source domain used for learning and the target domain where the system must work.

What it means

A model trained on one company's support emails may perform poorly on another company's messages. The task can still be ticket classification, but vocabulary, customers, products and writing styles have changed. Reusing the model without examining those differences risks confusing familiarity with capability.

The source domain supplies an existing model, data or both. The target domain is the intended new setting. Adaptation methods try to reduce the effect of the mismatch, often when target labels are limited or expensive.

This is a more specific problem than the broad idea of transfer learning. It also differs from domain generalization, where a system is developed to handle new domains without first adapting to their data. Practical projects should state which situation they actually face.

For a business, the important question is not whether knowledge can be transferred in principle. It is whether the adapted system performs reliably on the work, users and error types that matter in its new setting.

How it works

The team first compares the source and target settings. Changes may involve input patterns, class proportions, labeling rules or combinations of these. A method suited to one kind of shift may be unsuitable for another.

Possible approaches include adjusting the importance of source examples, learning representations that work across domains, or fine-tuning with available target labels. Some methods use unlabeled target data, but they still depend on assumptions about what information transfers.

The adaptation process is separated from evaluation. Target records used to tune a method should not also provide the sole evidence of its performance. A representative held-out target sample helps reveal whether the transfer helped.

The team also checks for negative transfer: reusing source information can make performance worse. A simple target-only baseline, where feasible, gives a useful comparison rather than assuming reuse must be beneficial.

Three practical examples

Customer service: adapt a classifier from one product line to another whose customers describe similar issues with different terminology.
Quality inspection: examine whether an image model trained under one factory's lighting can be adapted to a second production line.
Document processing: adapt extraction from familiar invoice layouts to a new supplier population while reviewing unfamiliar formats separately.

Think of it this way

Think of an experienced manager joining a new company. General skills transfer, but local procedures, customer expectations and the meaning of familiar labels may differ. Effective preparation combines existing knowledge with learning the new setting.

The analogy has a limit. A person can notice an unexpected organisational difference and ask about it. A model may confidently apply an old pattern unless its training and evaluation expose the mismatch.

Worked example

Checking the target rather than celebrating the source

Suppose a fictional ticket classifier correctly routes 940 of 1,000 held-out messages from its original business. On 200 reviewed messages from a new business, it routes only 150 correctly.

The source accuracy is 94%; the target accuracy is 75%. The first number does not describe the second deployment. The team investigates product-specific language and obtains a separate target development set for adaptation.

It then evaluates the adapted candidate on a fresh target test set, including rare escalation cases. The original 200 messages can support diagnosis, but repeated tuning against them would make them a weaker final test.

The figures are invented. They illustrate why a transfer project needs target evidence and a clear separation between development examples and final assessment. There is no promised improvement from adaptation itself.

Illustrative case study

A fictional distributor acquires another business and reuses its existing invoice classifier. Staff discover that the acquired company's document labels describe workflow stages rather than the document types used by the original system.

The technical team initially proposes matching visual styles. An editorial and operations review instead finds that the label definitions must be reconciled first. The team establishes a common task, then tests whether adaptation is still useful.

The lesson is that distribution differences are not always cosmetic. Reusing a model cannot resolve a disagreement about what its outputs are supposed to mean.

Common misconceptions and mistakes

A high source score proves readiness elsewhere. Only representative target evaluation supports the new deployment.
Unlabeled adaptation needs no assumptions. It still relies on a relationship between source and target information.
Any reuse must save money. Data review, adaptation, testing and maintenance can exceed the cost of a simpler alternative.

Business use

Domain adaptation is relevant when extending an AI system across locations, product lines, customer groups or document collections. It is particularly useful to discuss before a seemingly minor rollout becomes a performance problem.

Require the team to describe what changes, which target examples it has reviewed and how it will detect regressions. Compare the adapted model with practical alternatives using target outcomes and full operating cost.

Check your understanding

A vendor demonstrates excellent results on another company's documents but has not evaluated yours. What evidence is still missing?

Answer: Performance on representative target documents, including difficult and uncommon cases, evaluated separately from any adaptation or tuning examples.

Related terms

Learn next

Read Dataset shift and transfer learning. List differences between the training setting and the intended deployment before discussing a reuse strategy.

Sources

Kouw and Loog, An introduction to domain adaptation and transfer learning: https://arxiv.org/abs/1812.11806. Checked 12 October 2026. Business cases and numerical examples are original illustrations.

Previous: Domain Next: Domain closure