Domain Adaptation by Class Centroid Matching and Local Manifold Self-Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Domain adaptation is crucial for transferring knowledge between datasets with different distributions.
• Reducing distribution discrepancy is key for effective model learning across domains.
• Existing methods may not fully exploit the structural information within the target domain.

Purpose of the Study:

• To propose a novel domain adaptation approach that thoroughly explores target domain data distribution structure.
• To enhance knowledge transfer by treating clustered target samples as units and employing class centroid matching for pseudo-labeling.
• To leverage local manifold structure information for adaptive learning of target data's inherent connectivity.

Main Methods:

• A novel domain adaptation approach is proposed, focusing on clustering target domain samples.
• Pseudo-labels are assigned to target clusters via class centroid matching.
• A local manifold self-learning strategy is integrated to capture local sample connectivity.
• An efficient iterative optimization algorithm with convergence guarantees is developed.

Main Results:

• The proposed method effectively explores target domain data distribution and manifold structures.
• The approach demonstrates significant superiority in unsupervised domain adaptation.
• The method is successfully extended to semi-supervised domain adaptation (homogeneous and heterogeneous settings).
• Extensive experiments on seven benchmark datasets validate the proposed approach's effectiveness.

Conclusions:

• The novel domain adaptation approach offers a powerful way to reduce distribution discrepancies.
• The method's ability to explore target domain structure leads to improved learning.
• The approach provides a robust and versatile solution for both unsupervised and semi-supervised domain adaptation tasks.