Diffusion Graph Transformer for Learning Controllability Robustness in Large-Scale Networks

Area of Science:

• Complex Networks
• Network Science
• Systems Engineering

Background:

• Assessing network controllability robustness typically requires extensive simulations, which are computationally prohibitive for large-scale networks.
• Existing methods struggle with the efficiency and scalability needed for analyzing robustness against batch attacks.

Purpose of the Study:

• To develop an efficient and scalable method for learning network controllability robustness under batch attacks.
• To replace traditional, time-consuming simulation experiments with a data-driven approach.

Main Methods:

• Proposes the Diffusion Graph Transformer (DGT), a novel method leveraging graph attention mechanisms and diffusion principles.
• DGT generates node embeddings from degree attributes, propagates features via a diffusion strategy, and uses a fully connected layer for prediction.
• The approach is designed to handle batch attacks on complex networks.

Main Results:

• DGT achieves high accuracy and significant speed advantages over traditional simulation methods.
• Demonstrates strong generalization performance across various network scales, including networks with hundreds of thousands of nodes.
• Shows excellent transferability for predicting controllability robustness against different attack batch sizes.

Conclusions:

• DGT offers an effective and efficient solution for learning network controllability robustness, particularly for large-scale systems under batch attacks.
• The model's scalability and transferability make it a flexible tool for network resilience analysis.
• This work advances the field by addressing the previously computationally infeasible task of batch attack robustness learning on large networks.