A Graph Convolutional Network Approach for Predicting Network Controllability Robustness

Network controllability robustness reflects the ability of a network system to maintain its controllability against various attack strategies, which can be measured by a sequence of values that record the controllability of the remaining network after a sequence of node or edge-removal attacks. Convolutional neural networks can be used as a tool for predicting network controllability robustness, whose input is a gray-scale image converted from a network topology and the model size and the number of parameters are quite huge. In this paper, a graph convolutional network approch is developed for network controllability robustness prediction, in which a graph data along with its node characteristics is directly used as input without being converted to a gray-scale image. Experimental studies are carried out, which demonstrate that the proposed approach can obtain similar performance while the model size and the number of parameters have a hundredfold decline compared with the existing convolutional neural network approach.