The explosion of social media activity has extended the reach and impact of cyberbullying, presenting significant psychological and social risks. Effective moderation is challenged by platform-specific slang, the use of mixed media, complex context, and high user volume. This research introduces a cross-platform system for cyberbullying detection that leverages text, images, and conversational context using graph-based neural networks. The work includes the creation and annotation of a diverse, fine-grained multi-platform dataset. The proposed system utilizes a hybrid model integrating transformers, visual encoders, and relationship-focused neural networks. Results demonstrate improvements in generalization across platforms, multimodal understanding, and ethical fairness, enabling responsible and scalable deployment.
Sehrawat et al. (Tue,) studied this question.