Heterogeneous Domain Adaptation: An Unsupervised Approach.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Domain adaptation aims to improve learning in a target domain using knowledge from a source domain.
• Existing methods often require labeled data in the target domain, limiting their applicability.
• Heterogeneous unsupervised domain adaptation (HeUDA) with unlabeled target domains remains an under-researched area.

Purpose of the Study:

• To address the challenge of heterogeneous unsupervised domain adaptation (HeUDA).
• To propose a novel unsupervised knowledge transfer theorem and a domain distance metric.
• To develop an innovative transfer model, the Grassmann-linear monotonic maps-geodesic flow kernel (GLG), for HeUDA.

Main Methods:

• Developed an unsupervised knowledge transfer theorem ensuring correctness of knowledge transfer.
• Introduced a principal angle-based metric to quantify domain pair distances.
• Implemented the theorem and metric in the GLG model, using linear monotonic maps (LMMs) for homogeneous representations and geodesic flow kernel (GFK) for knowledge transfer.

Main Results:

• The GLG model successfully learns homogeneous representations of heterogeneous domains.
• Knowledge transfer is effectively achieved through these learned representations.
• Experiments on cancer detection, credit assessment, and text classification datasets showed superior performance compared to existing methods.

Conclusions:

• The proposed GLG model provides a significant advancement in heterogeneous unsupervised domain adaptation.
• The developed theoretical framework and metric offer a robust approach for knowledge transfer in unlabeled target domains.
• The model demonstrates broad applicability and effectiveness across diverse real-world applications.