Related Experiment Videos
PIMPC-GNN: Physics-Informed Multiphase Consensus Learning for Enhancing Imbalanced Node Classification in Graph
Summary
Physics-informed Graph Neural Networks (GNNs) improve imbalanced node classification by integrating diffusion, synchronization, and spectral embedding. PIMPC-GNN enhances minority class performance and offers interpretable insights.
Area of Science:
- Graph Neural Networks
- Machine Learning
- Network Science
Background:
- Graph neural networks (GNNs) face challenges in class-imbalanced scenarios, leading to biased predictions favoring majority classes.
- Underrepresented minority classes in graph data hinder accurate node classification and analysis.
Purpose of the Study:
- To introduce PIMPC-GNN, a novel physics-informed multiphase consensus framework designed to address class imbalance in node classification.
- To enhance the performance of GNNs on imbalanced graph datasets, particularly for minority classes.
Main Methods:
- Integration of three complementary dynamics: thermodynamic diffusion for label propagation, Kuramoto synchronization for minority node alignment, and spectral embedding for structural regularization.
- Class-adaptive ensemble weighting and an imbalance-aware loss function combining balanced cross-entropy with physics-based constraints.
Main Results:
- PIMPC-GNN demonstrated superior performance across five benchmark datasets with imbalance ratios ranging from 5 to 100.
- Achieved significant improvements over 14 state-of-the-art baselines, especially in minority-class recall and balanced accuracy.
- Provided interpretable insights into consensus dynamics within graph learning.
Conclusions:
- The proposed PIMPC-GNN framework effectively tackles class imbalance in node classification tasks.
- Physics-informed dynamics offer a promising direction for improving GNN robustness and interpretability in real-world applications.