Evaluating graph neural networks under graph sampling scenarios.

Area of Science:

• Network Science
• Machine Learning
• Graph Theory

Background:

• Real-world network analysis often deals with incomplete structural data.
• Existing network analysis methods typically assume complete network structures.
• Key questions involve method performance on partial networks and the benefit of structure completion.

Purpose of the Study:

• To evaluate the robustness of graph neural networks (GNNs) on incomplete network structures.
• To compare the performance of different GNNs and sampling methods.
• To determine the utility of completing missing network structures for downstream tasks.

Main Methods:

• Treated incomplete networks as random samples from complete graphs.
• Systematically evaluated six state-of-the-art GNNs.
• Utilized four common graph sampling techniques for evaluation.

Main Results:

• GNNs are applicable to incomplete network structures, with simpler models sometimes outperforming complex ones.
• Completing sampled subgraphs generally improves downstream task performance.
• Structure completion effectiveness is dataset-dependent and requires careful evaluation.

Conclusions:

• GNNs offer a viable approach for analyzing partially observed networks.
• Network structure completion can be beneficial but is not universally effective.
• Dataset-specific evaluation is crucial when deciding on structure completion strategies.