TGSL: Trade-off graph structure learning via multifaceted graph information bottleneck

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Mining

Background:

• Graph neural networks (GNNs) excel at processing graph data for node classification.
• Observed graph structures in real-world data are often suboptimal, hindering GNN performance.
• Existing GNNs rely on direct message passing over observed structures.

Purpose of the Study:

• To address the performance degradation of GNNs caused by suboptimal graph structures.
• To propose a novel method, Trade-off Graph Structure Learning (TGSL), for learning effective graph structures.
• To enhance node classification accuracy and robustness in GNNs.

Main Methods:

• Empirical analysis demonstrating the impact of graph structures on GNN performance.
• Development of TGSL, guided by the Graph Information Bottleneck (GIB) principle and Mutual Information (MI).
• Integration of global feature and structure augmentation, followed by structure refinement and redefinition.
• Optimization using multifaceted GIB to balance empirical risk minimization and information preservation.

Main Results:

• TGSL learns minimal sufficient graph structures that minimize empirical risk while preserving essential information.
• The method demonstrates superior performance across various datasets under both clean and attacked conditions.
• TGSL exhibits significant robustness compared to state-of-the-art GNN baselines.

Conclusions:

• TGSL effectively learns optimal graph structures, enhancing GNN performance for node classification.
• The proposed method offers a robust solution for handling suboptimal graph structures in real-world applications.
• TGSL represents a significant advancement in learning-based graph structure optimization for GNNs.