Heterogeneous Domain Adaptation via Nonlinear Matrix Factorization.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision
• Natural Language Processing

Background:

• Heterogeneous Domain Adaptation (HDA) addresses learning challenges with source and target data featuring distinct feature types.
• Existing HDA methods, often using matrix completion/factorization, effectively capture shared information but have limitations.
• Key limitations include the need for numerous corresponding data instances and the inability to capture nonlinear correlations.

Purpose of the Study:

• To propose a novel semi-supervised matrix-factorization-based HDA method.
• To overcome the limitations of existing HDA approaches, specifically the requirement for corresponding data instances and linear correlation capture.
• To enable effective HDA with minimal labeled target domain data and without paired cross-domain instances.

Main Methods:

• A novel semi-supervised HDA algorithm based on matrix factorization within an approximated Reproducing Kernel Hilbert Space (RKHS).
• The method leverages RKHS to exploit nonlinear correlations between features and data instances.
• It requires only a small amount of labeled data in the target domain, eliminating the need for cross-domain corresponding instances.

Main Results:

• The proposed method demonstrates superior performance compared to state-of-the-art HDA approaches.
• Experimental validation on cross-domain text classification and object recognition tasks confirms its effectiveness.
• The algorithm successfully learns heterogeneous features for both source and target domains by capturing nonlinear relationships.

Conclusions:

• The developed semi-supervised HDA method offers a practical and effective solution for cross-domain learning.
• It significantly advances HDA by addressing data requirements and capturing complex, nonlinear feature correlations.
• The approach shows promise for real-world applications where labeled target data is scarce and domain features differ.