Graph Neural Networks for Graph Drawing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Graph Theory

Background:

• Traditional graph drawing techniques aim for aesthetically pleasing node-link layouts.
• Differentiable loss functions have enabled gradient descent for graph layout optimization.
• Existing methods often struggle with complex graph structures and unlabeled vertices.

Purpose of the Study:

• To propose a novel framework for Graph Neural Drawers (GNDs) using neural computation for graph mapping.
• To enable GNNs to be trained with any differentiable loss function common in graph drawing.
• To demonstrate the use of feedforward neural networks and supervision hints for guiding GNDs.

Main Methods:

• Developed a framework for Graph Neural Drawers (GNDs) based on Graph Neural Networks (GNNs).
• Utilized differentiable loss functions, including those computed by feedforward neural networks, to guide the learning process.
• Incorporated positional features to handle unlabeled vertices within the GNN framework.
• Constructed a proof-of-concept loss function specifically for minimizing edge crossings.

Main Results:

• Demonstrated that GNNs can be effectively trained using diverse loss functions for graph drawing tasks.
• Showcased the ability of GNDs to incorporate positional features for handling unlabeled vertices.
• Provided quantitative and qualitative comparisons of different GNN models within the proposed framework.
• Successfully implemented a loss function to minimize edge crossings in graph layouts.

Conclusions:

• The proposed GND framework offers a flexible and powerful approach to neural graph drawing.
• GNNs, enhanced with positional features and guided by appropriate loss functions, can produce efficient and complex graph maps.
• This work opens new avenues for applying deep learning to graph visualization and analysis.