Wasserstein Graph Neural Networks for Graphs With Missing Attributes.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Science

Background:

• Missing node attributes are a significant challenge in real-world graph data.
• Existing graph neural networks (GNNs) do not adequately handle incomplete attribute information.
• This limitation hinders effective representation learning and downstream task performance.

Purpose of the Study:

• To propose a novel framework, Wasserstein Graph Neural Network (WGNN), for node representation learning on graphs with missing attributes.
• To maximize the utility of observed attributes and account for uncertainty from missing data.
• To enhance representation expressiveness by incorporating distributional information.

Main Methods:

• Representing nodes as low-dimensional distributions via attribute matrix decomposition.
• Employing a unique message-passing schema to aggregate distributional information in Wasserstein space.
• Evaluating WGNN on node classification using synthetic and real-world datasets with missing attributes.

Main Results:

• WGNN demonstrates superior performance in node classification tasks compared to existing methods.
• The framework effectively handles graphs with varying degrees of missing attributes.
• WGNN shows applicability in value recovery and matrix completion tasks, particularly in user-item graphs.

Conclusions:

• WGNN offers a robust solution for representation learning in graphs with missing node attributes.
• The method successfully leverages incomplete information and accounts for data uncertainty.
• WGNN advances the capabilities of graph neural networks in practical, data-scarce scenarios.