Social network analysis for crime prediction under social computing and deep learning technology

Abstract

Criminal networks feature complex structures and high internal concealment. Social computing methods and deep learning provide a possibility to identify hidden key figures and predict criminal behaviors within them. To this end, a Key Person Identification Model Based on Account Association (KPI-AA) is proposed. This model combines local neighbor similarity and global edge betweenness, and uses graph neural network (GNN) to conduct in-depth characterization of key nodes in social networks. It is applied to reveal potential organizational cores in complex social structures and identify potential diffusion paths in criminal networks. Experimental results show the following: In the dimension of propagation dynamics, KPI-AA infects 34 nodes at 40-time steps in the Zachary network, and reaches 329.6 nodes at 150-time steps in the Harry Potter network. Both results are higher than those of baseline models. In terms of network robustness, after removing 70% of the nodes in the Zachary network, the relative connectivity of KPI-AA is only 0.087. Regarding ranking consistency, the Kendall’s tau coefficient on the Twitter dataset reaches 0.467. Computational efficiency analysis indicates that while maintaining performance advantages in propagation dynamics, network robustness, and ranking consistency, KPI-AA still exhibits excellent scalability and practical deployment feasibility. The above results indicate that KPI-AA has advantages in propagation speed, revealing network vulnerability, and ranking consistency. Therefore, the KPI-AA model is practical in identifying core members of criminal networks and predicting criminal behaviors, and can serve social security governance and criminal investigation applications.