Attention Regularized Laplace Graph for Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Graph embedding methods in domain adaptation (DA) utilize Laplace graphs to preserve data manifolds.
• Existing methods neglect sub-domain adaptation and comprehensive manifold learning across feature/label spaces, leading to potential negative transfer.

Purpose of the Study:

• To propose a novel DA method, Attention Regularized Laplace Graph-based Domain Adaptation (ARG-DA), to overcome limitations of current graph embedding DA approaches.
• To enhance class-aware DA by weighting sub-domain adaptation tasks and unify manifold learning across feature/label spaces.

Main Methods:

• The proposed ARG-DA method incorporates an Attention Regularized Laplace Graph to weight sub-domain adaptation tasks, enabling class-aware DA.
• A FEEL strategy is employed to dynamically unify manifold structure alignment across different feature/label spaces for comprehensive manifold learning.

Main Results:

• ARG-DA consistently outperforms state-of-the-art DA methods on 7 standard DA benchmarks.
• The method demonstrates effectiveness across 37 cross-domain image classification tasks, including object, face, and digit images.

Conclusions:

• The proposed ARG-DA method effectively addresses limitations in existing graph embedding DA techniques.
• ARG-DA achieves superior performance in cross-domain image classification, showing robustness, convergence, and sensitivity advantages.