Unsupervised Domain Adaptation for Extra Features in the Target Domain Using Optimal Transport

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Domain adaptation typically assumes identical feature spaces between source and target domains.
• Few methods address domain adaptation with differing feature dimensions, especially without target domain labels.
• This research focuses on scenarios where the target domain has more features than the source domain.

Purpose of the Study:

• To develop a domain adaptation method for heterogeneous feature spaces.
• To address the challenge of knowledge transfer when target domain data has additional features.
• To provide a theoretical guarantee for the proposed method's performance.

Main Methods:

• Formulating domain adaptation as an optimal transport (OT) problem.
• Leveraging common features between source and target domains.
• Deriving a learning bound for the proposed OT-based method in the target domain.

Main Results:

• The proposed optimal transport-based domain adaptation method effectively handles differing feature dimensions.
• The method demonstrates successful knowledge transfer from source to target domains with new features.
• Validation on simulated and real-world data confirms the algorithm's efficacy.

Conclusions:

• The developed optimal transport approach offers a robust solution for domain adaptation with feature augmentation.
• This method advances the field by enabling adaptation in heterogeneous feature spaces without target labels.
• The theoretical learning bound provides confidence in the method's generalization capabilities.