Dynamic graph structure evolution for node classification with missing attributes.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Science

Background:

• Graph neural networks (GNNs) excel in many areas, but incomplete node attributes hinder performance.
• Existing graph completion learning (GCL) methods rely on accurate graph structures, which are often flawed.
• This limitation impacts the reliability of reconstructing missing node attributes.

Purpose of the Study:

• To propose the evolving graph structure (EGS) framework for semi-supervised node classification with missing attributes.
• To dynamically reconstruct node attributes and update graph structure simultaneously.
• To enhance the accuracy and robustness of GCL methods.

Main Methods:

• Introduced the evolving graph structure (EGS) framework.
• Employed an alternating optimization approach for dynamic attribute and structure reconstruction.
• Formulated an objective function using a Dirichlet Energy function with dual constraints.

Main Results:

• Demonstrated state-of-the-art performance on five benchmark datasets.
• Showcased effectiveness across various missing data rates.
• Validated EGS with seven different GNN variants, outperforming existing GCL methods.

Conclusions:

• The EGS framework effectively addresses challenges posed by incomplete node attributes in GNNs.
• Dynamic reconstruction of both attributes and graph structure is crucial for improved performance.
• EGS offers a robust solution for semi-supervised node classification in real-world scenarios with imperfect data.