Supervised Domain Adaptation: A Graph Embedding Perspective and a Rectified Experimental Protocol.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Domain Adaptation (DA) addresses data distribution discrepancies between source and target domains.
• Existing DA methods often rely on pairwise data relationships, which can be implicitly modeled using graph structures.

Purpose of the Study:

• To demonstrate that existing Supervised Domain Adaptation (SDA) methods can be formulated as Graph Embedding.
• To identify and address generalization and reproducibility issues in current SDA evaluation protocols.
• To propose a rectified evaluation protocol and establish updated benchmarks for SDA methods.

Main Methods:

• Analysis of loss functions in state-of-the-art SDA methods to reveal their Graph Embedding formulation.
• Identification of experimental setup limitations affecting few-shot learning demonstrations.
• Development of a revised evaluation protocol for SDA.

Main Results:

• Three leading SDA methods were shown to perform Graph Embedding by incorporating domain labels into graph structures.
• Generalization and reproducibility concerns were identified in common experimental practices.
• Updated benchmarks were generated for standard datasets including Office31, Digits, and VisDA.

Conclusions:

• The formulation of SDA as Graph Embedding offers a new perspective on existing methods.
• A standardized and rectified evaluation protocol is crucial for accurate SDA method comparison.
• The proposed benchmarks provide a reliable basis for future SDA research.