MuDiS-GDA: Multiscale discriminative graph domain adaptation.

Area of Science:

• Graph Domain Adaptation (GDA)
• Machine Learning
• Graph Representation Learning

Background:

• Graph Domain Adaptation (GDA) aims to transfer knowledge from labeled source graphs to unlabeled target graphs.
• Existing GDA methods often use a single adaptation strategy, ignoring scale-specific domain shifts.
• Unified strategies can lead to structural misalignment and negative knowledge transfer.

Purpose of the Study:

• To propose a multiscale GDA method that learns and aligns discriminative features across different graph structure scales.
• To address the limitations of unified adaptation strategies in current GDA approaches.
• To improve the accuracy of knowledge transfer and classification in target domains with limited labels.

Main Methods:

• Developed a multiscale contrastive learning framework with node-subgraph and node-graph contrasts to enhance feature discriminability.
• Implemented scale-specific domain adaptation strategies: adversarial adaptation for node-level features and bidirectional matching loss for subgraph-level alignment.
• Utilized source domain labels to create class prototypes for subgraph-level adaptation.

Main Results:

• The proposed multiscale GDA method effectively learns and aligns discriminative features at different structural scales.
• Scale-specific adaptation strategies reduce structural misalignment and negative knowledge transfer.
• Experimental results on real-world datasets demonstrate the superiority of the proposed method over strong baselines.

Conclusions:

• The multiscale GDA approach offers a more effective way to handle domain shifts in graph data.
• Adapting strategies to different structural scales enhances knowledge transfer and classification performance.
• The method successfully refines discriminative information from multiscale structures for improved GDA.