Link Prediction of UAV Networks Based on Dynamic Graph Neural Network

Link prediction is a topic within the realm of graph theory. However, the majority of existing link prediction models are tailored for static graphs, disregarding the temporal evolution of graphs and the significance of global features during this process. Addressing the high dynamics and time‐varying feature of unmanned aerial vehicle (UAV) network, traditional static graph methods encounter issues with the loss of network feature extraction information. By slicing the motion process of UAV nodes to form dynamic sequence graphs and designing sub‐modules of structural attention and temporal attention, we characterize the spatial relationships within subgraphs of dynamic sequence graphs and the temporal relationships between subgraphs. We propose a dynamic link prediction method using dynamic graph neural networks, which possesses exchangeability and interpretability, capturing the temporal dimension features of the entire evolution process of the cluster dynamic network and characterizing the spatial relationships between nodes. A dataset of time‐varying topologies for cluster networks is constructed, enabling link prediction under complex dynamic networks with time‐varying conditions. Experimental results demonstrate that this method can accurately predict link relationships between nodes in time‐varying complex UAV network topologies. Compared with traditional dynamic graph network models such as node2vec and GraphSAGE, this model achieves a link prediction accuracy of 88.79% on the RPGM dataset, surpassing node2vec's 70.24% and GraphSAGE's 62.81%.