Discriminative Transfer Subspace Learning via Low-Rank and Sparse Representation.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain transfer learning is crucial when target domain labels are unavailable.
• Existing methods struggle to effectively reduce domain discrepancy and preserve data structures.

Purpose of the Study:

• To develop a robust unsupervised domain transfer learning method.
• To reduce the discrepancy between source and target domains without target labels.
• To improve performance on visual domain adaptation tasks.

Main Methods:

• Data is transferred to a common subspace using a transformation matrix.
• Joint low-rank and sparse constraints are imposed on the reconstruction coefficient matrix.
• A non-negative label relaxation matrix is learned for a flexible linear classifier.
• The problem is formulated as a constrained minimization problem solved via the inexact augmented Lagrange multiplier method.

Main Results:

• The proposed method effectively reduces domain discrepancy, allowing source and target samples to interlace.
• Joint low-rank and sparse constraints preserve both global and local data structures.
• The method demonstrates robustness to noise by modeling it with a sparse matrix.
• Extensive experiments confirm the superiority over state-of-the-art methods in visual domain adaptation.

Conclusions:

• The developed method offers a superior approach to unsupervised domain transfer learning.
• It effectively addresses the challenge of label scarcity in target domains.
• The technique shows significant improvements in visual domain adaptation tasks.