DREAM: A Dual Variational Framework for Unsupervised Graph Domain Adaptation.

Area of Science:

• Graph Machine Learning
• Artificial Intelligence
• Data Mining

Background:

• Graph classification is crucial in machine learning, often using Message Passing Neural Networks (MPNNs).
• MPNNs implicitly learn topological semantics but struggle with domain shift and limited labels in unsupervised domain adaptation.
• Existing methods face challenges in extracting comprehensive graph structural semantics.

Purpose of the Study:

• To propose Dual Variational Semantics Graph Mining (DREAM) for unsupervised graph domain adaptation.
• To combine implicit and explicit graph structural semantics for improved graph representations.
• To enhance the robustness and accuracy of graph classification models in domain adaptation scenarios.

Main Methods:

• DREAM employs a dual-branch architecture: a message passing branch for implicit semantics and a path aggregation branch for explicit high-order structural semantics.
• An expectation-maximization (EM) style variational framework is utilized for conjoint training of both branches.
• The E-step constructs a graph-of-graph to capture inter-domain correlations, while the M-step refines the MPNNs.

Main Results:

• The proposed DREAM method demonstrates superior performance compared to existing baselines on benchmark datasets.
• The joint optimization of implicit and explicit semantics effectively improves graph classification.
• Experimental validation confirms the efficacy of the DREAM approach in unsupervised domain adaptation.

Conclusions:

• DREAM offers an effective solution for unsupervised graph domain adaptation by leveraging complementary graph structural semantics.
• The integration of explicit and implicit learning pathways enhances the generalization capabilities of graph neural networks.
• The study validates the superiority of DREAM, paving the way for more robust graph classification models.