Identification of Critical Subgraphs in Drone Airways Graphs by Graph Convolutional Networks

Abstract

Our original aim was to study the control of drones in a congested airspace that may be structured in airways much similar to road or subway traffic networks. The question posed was if it is possible to identify subgraphs whose congestion that may lead to the blockade of the entire network, i.e. if there are dominant subgraphs in the network. We resort to semi-supervised trained Graph Convolutional Networks (GCNs) to formulate and solve the problem. For traffic networks structured as streets or subway lines, the notion of subgraph dominance is very intuitive, because we can postulate that if the labeling of the graph assigns a line label to many nodes belonging to other lines, then the flow in this line should affect these nodes and their corresponding lines. We have carried out experimental work over a manageable network, the Vienna subway network, achieving high and robust accuracy results in the semi-supervised labeling process. However, line dominance results are less robust, they seem to be highly influenced by the experimental setting given by the actual seed nodes provided to networks, which calls for further experiments and analysis in order to try to find stable responses.