Feature Space Independent Semi-Supervised Domain Adaptation via Kernel Matching

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Domain adaptation (DA) aims to leverage labeled source data for prediction in label-scarce target domains.
• Significant feature distribution divergence and heterogeneous feature representations pose challenges in practical DA scenarios.
• Existing methods struggle when source and target domain features differ substantially.

Purpose of the Study:

• To propose a novel feature space independent semi-supervised kernel matching method for domain adaptation.
• To address substantial feature distribution divergence and heterogeneous feature representations across domains.
• To enable effective learning in label-scarce target domains by mapping target data to similar source data.

Main Methods:

• Developed a kernel matching method that learns a prediction function on source data while mapping target data.
• Utilized the Hilbert Schmidt Independence Criterion to match the target kernel matrix to a submatrix of the source kernel matrix.
• Formulated the process as a non-convex integer optimization problem, solved via a relaxed continuous form and local minimization.

Main Results:

• The proposed kernel matching method demonstrated consistent and significant performance improvements over comparison methods.
• Effectiveness was validated on cross-domain sentiment classification (Amazon product reviews) and cross-language text classification (Reuters newswire stories).
• The method excelled in both homogeneous and heterogeneous feature space scenarios, showcasing its robustness.

Conclusions:

• The proposed feature space independent semi-supervised kernel matching method is effective for domain adaptation.
• It successfully handles significant feature distribution divergence and heterogeneous feature representations.
• The approach offers a robust solution for improving prediction models in label-scarce target domains.