<p>The proliferation of online social networks has amplified the risks of cybercrime, including cyberbullying, terrorism-related activities, and online threats. Addressing these challenges requires intelligent systems capable of understanding not only textual content but also the underlying network interactions to ensure platform safety and reliability. This study explores a suite of machine learning (ML) and deep learning (DL) models for cybercrime detection, evaluating their performance using real-world social media data. In this paper, alongside traditional architectures, we propose a new model – Transformer-Augmented Graph Neural Network (T-GNN), which synergistically integrates a transformer’s contextual comprehension with a graph neural network’s structural understanding. The T-GNN integrates rich semantic feature extraction from text with user interaction modeling through graphs, capturing the essence of social media behavior. Under varying train-test splits and multiple feature extraction strategies, comparative experiments demonstrate that while DL models surpass traditional ML methods, T-GNN continues to achieve the highest accuracy and robustness for all tasks. This research highlights the role of sophisticated models to transform the trustworthiness and safety of digital systems and contribute towards intelligent cybercrime detection systems.</p>