Hybrid GNN-Based Link Prediction Model for Identifying Drug-Related Organized Crime Groups on Twitter

In this paper, we propose a Graph Neural Network (GNN)-based link prediction model to analyze the interconnections of Drug-related Organized Crime Groups (DOCG) on X (formerly Twitter) and to identify their potential associations. To this end, we collected approximately one million tweets by employing slang terms commonly used for drug distribution and promotion. From these data, we constructed a knowledge graph in which tweets containing URLs to secure messengers such as Telegram and Snapchat, the accounts posting these tweets, and the hashtags included in them were modeled as nodes, thereby enabling the identification of accounts belonging to the same DOCG. In particular, to compare the impact of distinguishing node types (tweets, accounts, hashtags) on link prediction performance, we generated both homogeneous and heterogeneous knowledge graphs. For link prediction, we employed three GNN architectures: Graph Convolutional Network (GCN), Graph Attention Network (GAT), and Graph Isomorphism Network (GIN). We further developed single-model frameworks as well as hybrid models combining two architectures, with the objective of determining the most effective GNN for DOCG identification. Experimental results demonstrate that, across both homogeneous and heterogeneous settings, the GAT model achieved the highest F1-score among the single models, while the heterogeneous hybrid model combining GAT and GIN yielded the best overall performance. These findings indicate that GNN-based link prediction is effective in detecting latent associations among DOCG on social media platforms.