| Complex network is one of the hot research fields at present.The interdisciplinary research between network science and control system theory has developed rapidly in the past 20 years.Complex networks explore the connection relationship between edges and nodes at the micro level and explain the macroscopic system characteristics of the network,which is crucial to the study of the characteristics of large-scale complex systems in reality.Complex systems exist widely in the Internet,transportation,neural,genetic and other fields,and can also be used to study the interaction of interpersonal social relations,academic cooperation,human migration and other interactions.Complex networks have been widely used in mathematics,engineering,economics and other disciplines,and are closely related to our daily life,providing valuable reference models.The robustness of a complex network refers to the ability of the network to maintain its original structure and function in the presence of deliberate attacks or random failures.Predicting the robustness of the network can help us evaluate the stability,reliability and security of the network system,and provide an important reference for network optimization and design.In order to evaluate the robustness of complex networks,it is necessary to design specific attack strategies and observe the connectivity or robustness changes of the network during the whole attack process.However,the time complexity of calculating the controllability of a network is 0(N2.37),which is closely related to the size of the network,without considering the time consuming of the attack strategy.In addition,because complex networks usually contain a large number of nodes and connected edges,the computation of robustness analysis is too inefficient.Especially for large scale real network systems,the evaluation of network robustness becomes particularly difficult.Therefore,evaluating network robustness in large-scale network systems has always been a hot and difficult topic in academia and industry,and more efficient and feasible algorithms and techniques have been explored to evaluate network robustness quickly and accurately.Aiming at the pain point of huge calculation of robustness evaluation,this project intends to explore different deep learning algorithms,build efficient and high-precision offline robustness predictors driven by data,and complete the evaluation and calculation research of complex network robustness.The main work of this paper will be composed of three parts,which are 1)complex network robustness predictor built by single CNN architecture.In this part of work,for the task of robustness prediction,the CNN structure and the illegal numerical correction filter based on prior knowledge are designed to predict the connectivity robustness,and the time of robustness calculation is greatly reduced on the basis of ensuring high prediction accuracy,which provides a preliminary idea for using convolutional neural network for network robustness prediction.2)A robust predictor of multi-CNN architecture with prior knowledge of network topology.This part of the work is the extension and expansion of the first part.The prior knowledge of the network topology type is added,and the end-to-end prediction is performed on the basis of the topology classification,which further improves the accuracy of the robustness prediction compared with the work in the first part.3)Graph representation learning based robustness predictor for complex networks.This part of the work aims at designing robust predictors with better generalization.Aiming at the pain point that the first two predictors cannot deal with networks of variable size,this part of the work adds a graph representation learning operator to extract and rank the key node features of the network,and uses a lightweight CNN structure for regression prediction.The proposed predictor greatly improves the accuracy of robust prediction under varying range of input scenarios,as well as the generalization ability of robust predictor. |
PDF Full Text Request