Adaptive Betweenness Clustering for Semi-Supervised Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Semi-supervised domain adaptation (SSDA) utilizes scarce labeled target data to boost model performance and generalization.
• Existing SSDA methods struggle with semantic alignment due to limited target domain label information.

Purpose of the Study:

• To introduce a novel SSDA approach, Graph-based Adaptive Betweenness Clustering (G-ABC), for effective categorical domain alignment.
• To enable cross-domain semantic alignment by transferring knowledge from labeled source and target data to unlabeled target samples.

Main Methods:

• Construct a heterogeneous graph representing relationships between labeled and unlabeled samples across domains.
• Refine graph connectivity using Confidence Uncertainty based Node Removal and Prediction Dissimilarity based Edge Pruning.
• Employ Adaptive Betweenness Clustering for semantic transfer via across-domain and within-domain clustering.

Main Results:

• The proposed G-ABC method achieved superior performance compared to state-of-the-art SSDA approaches.
• Experiments on DomainNet, Office-Home, and Office-31 datasets validated the effectiveness of G-ABC.

Conclusions:

• G-ABC successfully addresses the challenge of semantic transfer in SSDA.
• The method demonstrates significant improvements in classification performance and generalization capability.