Heterogeneous Domain Adaptation via Covariance Structured Feature Translators

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Domain adaptation (DA) is crucial for image analysis and data classification, especially with heterogeneous data where features differ significantly.
• Existing methods struggle with heterogeneous data due to differing feature dimensions and sample distributions, hindering direct feature matching.

Purpose of the Study:

• To address the challenge of domain adaptive feature representation for heterogeneous data.
• To develop a method for efficiently transferring discriminant information from source to target domains to improve target data classification.

Main Methods:

• Proposed a joint kernel regression model to learn a 'feature translator' using two domain-specific projection matrices.
• Introduced optimal experimental design (OED) to find covariance structured feature translators (CSFTs) for nonlinear DA.
• Developed an efficient method for computing optimal data projections.

Main Results:

• The proposed method demonstrates state-of-the-art performance in heterogeneous domain adaptation.
• Comprehensive experiments validated the effectiveness and efficacy of the CSFTs approach.
• Achieved superior classification performance on target data by effectively bridging domain gaps.

Conclusions:

• The novel CSFTs approach provides an effective solution for heterogeneous and nonlinear domain adaptation.
• The method successfully transforms heterogeneous features into a shared space for improved transfer learning.
• This research advances the field of domain adaptation with a robust and efficient technique.