Session-Level Cyberbullying Detection using Neural Networks

Abstract

The rise of social media platforms has led to an increase in cyberbullying incidents, necessitating the development of automated detection systems. This research focuses on session-level cyberbullying detection using neural networks, leveraging deep learning techniques to identify harmful online interactions. We explore various machine learning and deep learning models, including Temporal Convolutional Networks (TCNs) and Gated Recurrent Units (GRUs), to enhance detection accuracy. The study also incorporates transformer-based models, particularly multi-head self-attention mechanisms, to improve contextual understanding in textual data. Our methodology involves data preprocessing, feature extraction, and model training using real-world cyberbullying datasets from Instagram and Vine. The performance of the proposed models is evaluated using standard metrics such as precision, recall, F1-score, and Area Under the Curve (AUC). Experimental results demonstrate that context-aware models significantly outperform traditional post-level approaches, especially when incorporating positional encoding and self-attention layers. For the Instagram dataset, the TCN model with 60-comment sequences achieved the highest recall of 88% for the cyberbullying class, highlighting its strong sensitivity to harmful interactions. On the Vine dataset, the TCN model with 100 comments reached the best recall of 84% for the cyberbullying class. These findings emphasize the importance of recall in detecting harmful content and the impact of architecture design and context length on model performance. This study contributes to the advancement of cyberbullying detection systems by providing a robust framework for automated content moderation. Future research will focus on improving model generalization across diverse online platforms and enhancing interpretability for better real-world application.