Visualizing Graph Neural Networks With CorGIE: Corresponding a Graph to Its Embedding.

Area of Science:

• Computer Science
• Machine Learning
• Data Visualization

Background:

• Graph neural networks (GNNs) are advanced machine learning models for analyzing relational data.
• Evaluating GNNs often relies on quantitative metrics, making it challenging to ascertain if they capture underlying graph structures.
• Understanding the internal representations (node embeddings) learned by GNNs is crucial for effective model development.

Purpose of the Study:

• To develop a method for correlating input graphs with their corresponding node embeddings in GNNs.
• To create an interactive visualization tool, CorGIE, to aid GNN developers in understanding model behavior.
• To introduce a novel K-hop graph layout for visualizing topological neighborhoods and clustering within graph data.

Main Methods:

• Developed an interactive multi-view interface named CorGIE.
• Implemented a K-hop graph layout as a core visualization component within CorGIE.
• Designed CorGIE to abstract graph data and tasks for intuitive exploration.
• Evaluated CorGIE through usage scenarios and a case study with GNN experts.

Main Results:

• CorGIE successfully instantiates the abstraction of graph data and tasks.
• The K-hop graph layout effectively visualizes topological neighbors and their clustering.
• A case study with GNN experts demonstrated the functionality and usability of CorGIE.
• The tool facilitates a deeper understanding of how GNNs learn graph characteristics.

Conclusions:

• CorGIE provides a valuable tool for GNN developers to interpret model learning.
• The interactive visualization approach enhances the understanding of GNN embeddings.
• CorGIE, with its K-hop layout, addresses the challenge of evaluating GNNs beyond quantitative metrics.