Adaptive Graph Learning with Semantic Promotability for Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Domain Adaptation (DA) aims to bridge the gap between labeled source and unlabeled target domains.
• Existing semantic-based DA methods often struggle with personalized intra-class and local inter-class knowledge acquisition.
• This limitation hinders comprehensive knowledge transfer across domains.

Purpose of the Study:

• To propose a novel Domain Adaptation approach, Adaptive Graph Learning with Semantic Promotability (AGLSP).
• To address limitations in acquiring personalized and local knowledge in cross-domain settings.
• To enhance the transfer of semantic knowledge between source and target domains.

Main Methods:

• AGLSP employs Adaptive Graph Embedding with Semantic Guidance (AGE-SG) to estimate target sample promotability and learn domain-specific components.
• Semantically Promotable Sample Enhancement (SPSE) refines feature discriminability using multi-granular intra-class and target sample mining.
• Adaptive Graph Learning with Implicit Semantic Preservation (AGL-ISP) extracts commonalities for tag granularity, even from non-promotable target samples.

Main Results:

• The proposed AGLSP method demonstrates superior performance in domain adaptation tasks.
• Extensive experiments on seven datasets validate the effectiveness of the AGLSP approach.
• AGLSP successfully transfers more cross-domain knowledge by learning richer semantic information.

Conclusions:

• AGLSP offers a robust framework for Domain Adaptation by addressing limitations of existing methods.
• The multi-semantic-granularity and target-sample oriented strategy significantly improves knowledge transfer.
• The approach shows strong potential for various cross-domain learning applications.