Learning on Attribute-Missing Graphs.

Area of Science:

• Graph Machine Learning
• Network Science
• Data Science

Background:

• Real-world graphs often have missing node attributes, limiting the performance of existing graph learning methods.
• Current graph neural networks (GNNs) are not optimized for handling attribute-missing graph data.
• Developing specialized GNNs for attribute-missing graphs is crucial for advancing graph learning.

Purpose of the Study:

• To develop a novel GNN model capable of learning from graphs with incomplete node attributes.
• To address the challenges posed by attribute-missing graphs in real-world applications.
• To introduce and evaluate a new node attribute completion task.

Main Methods:

• Proposed a Structure-Attribute Transformer (SAT), a distribution matching-based GNN for attribute-missing graphs.
• Leveraged a shared-latent space assumption and decoupled scheme for structures and attributes.
• Employed distribution matching techniques for joint modeling of graph structures and attributes.

Main Results:

• SAT demonstrated superior performance on both link prediction and the novel node attribute completion task.
• Experiments on seven real-world datasets validated the effectiveness of SAT.
• Introduced practical metrics for quantifying node attribute completion performance.

Conclusions:

• SAT offers a significant advancement in learning from attribute-missing graphs.
• The model successfully handles incomplete data, improving performance on key graph learning tasks.
• SAT provides a robust framework for future research in graph learning with incomplete information.