IMMENSE: Inductive Multi-perspective User Classification in Social Networks

Online social networks are increasingly used to propagate hate speech, radical ideologies, and violent content. Law Enforcement Agencies (LEAs) require automated tools to identify these malicious users at scale. Existing approaches often rely on a single perspective—such as text analysis or network topology—or operate in a transductive setting, which requires the entire network (including new users) to be present during training.

To address these limitations, the authors introduce IMMENSE (Inductive Multi-perspective Model for usEr classificatioN in Social nEtworks). IMMENSE uses a hybrid strategy that simultaneously analyzes three distinct modalities: the semantics of user-generated content, social relationships (e.g., follows, retweets), and spatial information (geographical proximity). By adopting an inductive learning approach, the model learns to generalize from labeled training data to new, unseen users, making it suitable for real-world, dynamic social network monitoring.

The researchers evaluated IMMENSE against five state-of-the-art competitors using a real-world Twitter/X dataset. The results demonstrate that IMMENSE consistently outperforms single-modality and existing hybrid approaches. By integrating content, social, and spatial perspectives, the model captures complex patterns of radicalization that are often missed when analyzing these dimensions in isolation. Furthermore, the inductive nature of the framework allows for immediate classification of new users, significantly reducing the computational overhead and maintenance requirements compared to transductive models that necessitate frequent retraining.

This work provides a scalable and sustainable solution for identifying malicious actors in social networks. By enabling the classification of new users without retraining, IMMENSE offers a practical tool for LEAs to monitor evolving social environments in real time. The study highlights the importance of multi-perspective analysis, suggesting that combining content semantics with structural and spatial data is essential for detecting sophisticated malicious behavior that might otherwise remain hidden.