Identifying Drug Users on Social Media

Lucrative profits have driven the evolution of illicit drug trafficking in recent decades, chiefly on social media platforms, which offer direct-to-consumer mediums for drug trade. The rapidly evolving landscape of social media enables sellers to adapt and avoid algorithmic detection, posing a significant threat to public health and safety. Previous research mainly leverage (Heterogeneous) Graph Neural Networks to tackle this issue that merely integrate the content information, i.e., user profile and posts, into user features and consider the relationships among these user nodes, while ignoring the complex relationships between entities, i.e., users and posts. Moreover, most existing drug trafficking dataset are balanced data which fails to mimic the real-world scenario. In light of this, we propose an imbalance Heterogeneous Graph Neural Network for drug trafficking (iHG-DT) to  learn the user representations from both forms of heterogeneity (nodes and edges). To better mimic the complex relationships in real-world application on social media, we build a new imbalanced dataset from Twitter called Twitter-HetDrug that integrates three types of nodes and seven types of relationships. We benchmark our proposed model against (Heterogeneous) Graph Neural Networks on the Twitter-HetDrug and DBLP heterogeneous graph datasets. Our proposed method (iHG-DT) outperforms all baseline methods in F1 macro and accuracy, which demonstrates its superiority in learning from complex semantic heterogeneous relationships.