Critical Node Identification based on Attention Flow Networks

The identification of critical nodes in complex networks has been extensively studied, but little research has been done on the identification of critical nodes in attentional flow networks. Based on the massive online user behavior data provided by CNNIC, this paper proposed a key node identification model GAT-RL (Graph Attention Networks and Reinforcement Learning) by effectively using the node characteristics and the impact on the network after node removal. Firstly, a directed weighted attentional flow network was constructed based on online user behavior data. Then Graph Attention Networks (GATs) were used to aggregate the neighborhood features of each node in the network to obtain a vector representation of each node. Finally, the vector representation of each node is mapped to the corresponding quality score in combination with reinforcement learning. The critical node ranking is obtained based on the scores. Experiments show that when using the four methods of the GAT-RL model, H-index, degree centrality (Degree) and graph convolutional neural network (RCNN) to identify key nodes in the attention flow network, the GAT-RL model has the most rapid decrease in network connectivity during the identification process. When the node removal ratio is 4%, the connectivity of the remaining graph is about 0.367. When the node removal ratio reaches 13%, the connectivity of the remaining graph is close to 0. Therefore, the GAT-RL model can quickly and accurately identify the critical nodes in the attention flow network.