A hierarchical multimodal transformer fusion network for cyberbullying detection across social media platforms

Cyberbullying has become a critical challenge for online social media platforms due to its harmful psychological and social consequences and the increasing volume of user-generated content. Existing cyberbullying detection methods predominantly rely on textual information and often fail to effectively capture conversational context, user interaction structures, visual content, and behavioral metadata, resulting in limited robustness and interpretability. This paper proposes HMTF-Net, a Hierarchical Multimodal Transformer Fusion Network for explainable cyberbullying detection across social media platforms. The proposed framework jointly learns complementary representations from textual content, conversation threads, heterogeneous user interaction graphs, images, and metadata through dedicated representation learning branches. An adaptive cross-modal attention mechanism with gated fusion integrates heterogeneous modalities, while a hierarchical conversation encoder captures discourse-level dependencies and a graph attention encoder models structural relationships among interacting users. Furthermore, an explainability module provides interpretable predictions by highlighting influential words, conversation turns, graph nodes, and image regions, whereas a domain-adversarial learning strategy enhances cross-platform generalization by learning platform-invariant representations. Experimental evaluations on publicly available cyberbullying datasets demonstrate that HMTF-Net consistently outperforms representative state-of-the-art approaches, achieving 97.6% accuracy, 97.0% precision, 96.2% recall, 96.6% F 1-score, and a 0.986 ROC-AUC while maintaining competitive computational complexity and inference efficiency. Ablation studies confirm the contribution of each proposed component, and robustness experiments demonstrate strong generalization across multiple social media platforms and noisy linguistic conditions. These results indicate that HMTF-Net provides an accurate, interpretable, and scalable solution for trustworthy automated cyberbullying detection in real-world online environments.