Unsupervised domain adaptation is the category of domain adaptation that aims to train the model with labeled source domain adapt to the unlabeled target domain. Target domains are similar but have different data distributions. Unsupervised domain adaptation handles the domain shift problem by using the unlimited labeled source data that adapt to target data.

The significance of unsupervised domain adaptation is the ability to transfer knowledge learned from source domains with a large number of annotated training sets to target domains with unlabeled data only. Conventional approaches of unsupervised domain adaptation are based on matching the feature distribution between source and target domain. Methods of conventional approaches are sample re-weighting and feature space transformation.

Recently deep learning approaches are emerged to learn powerful features. Deep learning approaches utilize generative adversarial networks to align the features of different domains. Other new approaches of unsupervised domain adaptation based on neural networks are domain invariant feature learning, domain mapping, normalization statistics, ensemble methods, target discriminative methods, and combinations of these methods.

Unsupervised domain adaptation application areas are computer vision, natural language processing, time series, domain generalization. Future enhancements of unsupervised domain adaptation are bi-directional adaptation, hyper-parameter tuning, balancing classes, incorporating improved image to image translation methods, heterogeneous feature spaces handling, and multi-source or target domains data handling.