MagNet: A Neural Network for Directed Graphs.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Network Science

Background:

• Graph-based data is increasingly prevalent, driving advancements in machine learning algorithms.
• Existing research predominantly focuses on undirected graphs, neglecting the potential of directed graph structures.
• Directed graphs are common in real-world networks like citation, website, and traffic systems.

Purpose of the Study:

• To propose MagNet, a novel graph neural network (GNN) specifically designed for directed graphs.
• To leverage the magnetic Laplacian for encoding both geometric structure and directional information.
• To demonstrate the efficacy of MagNet on various directed graph tasks.

Main Methods:

• Developed MagNet, a GNN architecture utilizing a complex Hermitian matrix called the magnetic Laplacian.
• The magnetic Laplacian encodes undirected structure via magnitude and direction via phase.
• Introduced a 'charge' parameter to tune spectral information for directed cycles.

Main Results:

• MagNet was applied to node classification and link prediction tasks on directed graphs.
• The proposed network demonstrated strong performance across all evaluated tasks.
• MagNet outperformed existing methods on a majority of the directed graph tasks.

Conclusions:

• MagNet offers a powerful new approach for analyzing directed graph data.
• The magnetic Laplacian provides an effective mechanism for incorporating directional information into GNNs.
• The principles behind MagNet are adaptable to other GNN architectures, suggesting broad applicability.