Generic Structure Extraction with Bi-Level Optimization for Graph Structure Learning.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Artificial Intelligence

Background:

• Current Graph Structure Learning (GSL) methods often focus on local graph information, limiting Graph Neural Network (GNN) robustness.
• Local GSL approaches struggle with graph's local structure heterogeneity, where inter-class connections are unevenly distributed.
• This heterogeneity can hinder the performance of GNNs in real-world applications.

Purpose of the Study:

• To develop a novel GSL method that learns graph structure from a global perspective.
• To enhance GNN robustness by jointly optimizing graph structure and common GNN parameters.
• To address the limitations of local GSL methods in handling graph structure heterogeneity.

Main Methods:

• Proposed Generic Structure Extraction with Bi-level Optimization for Graph Structure Learning (GSEBO).
• Extracted graph structure as a learnable parameter, enabling joint learning from a global view.
• Modeled the learning process as a bi-level optimization: GNN parameters in the upper level, graph structure in the lower level.

Main Results:

• GSEBO effectively learns graph structure and common parameters from a global viewpoint.
• The common parameters provide crucial global mapping information for structure optimization.
• Experiments demonstrated GSEBO's effectiveness compared to state-of-the-art GSL methods on four real-world datasets.

Conclusions:

• GSEBO offers a significant advancement in GSL by incorporating global information.
• The bi-level optimization framework successfully enhances GNN robustness.
• The proposed method shows strong performance and generalizability across diverse datasets.