domainadaptation

Domain adaptation is a subfield of transfer learning concerned with adapting a model trained on one data distribution, the source domain, to perform well on a different distribution, the target domain. The distributions may differ in feature statistics, label distributions, or the relationship between features and labels. The objective is to improve target-domain performance when labeled data in the target domain is scarce or unavailable.

Common settings include supervised domain adaptation, where labeled target data is available; unsupervised domain adaptation, where target labels are not; semi-supervised domain adaptation; and multi-source domain adaptation, which uses multiple source domains. The key challenge is mitigating distribution shift between domains while preserving task performance.

Approaches fall into several families. Instance reweighting adjusts the contribution of source samples to better match the target. Representation learning aims to extract domain-invariant features. Discrepancy-based methods minimize distance measures between source and target feature distributions, such as maximum mean discrepancy or CORAL. Adversarial methods train a domain discriminator to encourage indistinguishable domains, as in domain-adversarial neural networks. Other strategies include self-training with pseudo-labels and fine-tuning or adapting models for the target domain.

Evaluation typically uses target-domain test sets and standard metrics like accuracy or AUC. Datasets such as Office-31, Office-Home, VisDA, and various digits datasets are common benchmarks in vision; NLP and speech domains employ corresponding corpora. Challenges include negative transfer when domains are highly dissimilar, covariate and label distribution shifts, and limited labeled target data. Ongoing work seeks stronger theoretical guarantees and more sample-efficient, robust methods across modalities.