Network Robustness Prediction: Influence of Training Data Distributions.

Area of Science:

• Network science
• Computational science
• Machine learning

Background:

• Network robustness is crucial for natural and industrial systems, traditionally assessed via time-consuming simulations.
• Evaluating network robustness involves measuring remaining functionality after node or edge removal attacks.
• Convolutional neural networks (CNNs) present a computationally efficient alternative for robustness prediction.

Purpose of the Study:

• To compare the prediction performance of learning feature representation-based CNN (LFR-CNN) and PATCHY-SAN for network robustness.
• To investigate the impact of different training data distributions (uniform, Gaussian, extra) on CNN performance.
• To analyze the relationship between CNN input size and network dimension for robustness prediction.

Main Methods:

• Empirical experiments comparing LFR-CNN and PATCHY-SAN prediction accuracy.
• Investigation of network size distributions: uniform, Gaussian, and extra.
• Analysis of CNN input size versus network dimension effects on robustness prediction.
• Evaluation of generalization and extension abilities on unseen networks.

Main Results:

• Gaussian and extra data distributions significantly enhance prediction performance and generalizability for both LFR-CNN and PATCHY-SAN.
• LFR-CNN demonstrates superior extension ability compared to PATCHY-SAN in predicting robustness of unseen networks.
• LFR-CNN generally outperforms PATCHY-SAN across various robustness measures.

Conclusions:

• LFR-CNN is recommended over PATCHY-SAN for network robustness prediction due to its superior performance and generalizability.
• The choice between LFR-CNN and PATCHY-SAN may depend on specific application scenarios.
• Optimal CNN input sizes are recommended for different configurations to maximize prediction accuracy.