Predicting the Robustness of Real-World Complex Networks

Abstract

Many real-world natural and social systems can be modeled as complex networks. Since random failures and malicious attacks can seriously destroy the structure of the complex networks, it is critical to ensure its robustness and maintain the functions. Generally, connectivity robustness and controllability robustness are adopted to evaluate the performance of the networked systems against external attacks and/or failures. A sequence of values are measured to dynamically indicate the network robustness with iterative node- or edge-removal. Calculating the robustness of large-scaled real-world networks is usually time consuming, while deep-learning has provided an efficient methodology to estimate network robustness performance. In this paper, a multi-convolutional neural network (CNN) method called Real-RP is designed for predicting the robustness of real-world complex networks. Unknown real-world networks are firstly classified into known network category, and their robustness performance is then predicted based on the knowledge of the specific network category that is trained using a substantial number of synthetic networks. Experimental results show that: 1) real-world complex networks can be classified by CNN with high precision; and 2) the robustness performance of real-world networks can be predicted with low average errors than the existing methods.