Prompt-Tuned Cross-Level Transformer for Behavioral Risk Detection: A Deep Learning Multi-Granular Approach to Mass Shooting Threat Prediction

Mass shootings in educational environments represent a growing public safety crisis, as traditional security strategies have proven insufficient for early prevention. Although many perpetrators exhibit early warning signs through their online behavior, these signals often go undetected. This research introduces a novel deep learning framework for the proactive detection of mass shooting threats by analyzing social media content in real time. The model is built around a prompt-tuned cross-level transformer architecture, enhanced with structured behavioral lexicons related to violence, radicalization, and psychological distress. The proposed system integrates multi-granular threat analysis, simultaneously capturing signals at the post level (linguistic cues and sentiment patterns) and the user level (temporal posting behavior and aggregated risk). Prompt-tuning is used to adapt the transformer to subtle behavioral contexts with minimal supervision, while domain-specific lexical knowledge dynamically influences the model’s attention and interpretability. This multi-layered design allows the system to prioritize high-risk language, understand intent, and adapt to varying linguistic expressions of violence. The dataset for this research will be collected from publicly accessible sources on Reddit and YouTube comment threads on videos discussing mass shootings or violent ideologies. These platforms are known to contain real-world behavioral signals, including linguistic cues of violent intent, emotional distress, and radicalization. All data will be anonymized and processed in accordance with ethical research standards. The framework is grounded in General Strain Theory, which links emotional distress and perceived injustice to deviant behavior, and Routine Activity Theory, which contextualizes online behavior within the conditions that enable harmful actions. These theories support the model’s design, enabling it to move beyond surface-level text mining toward behaviorally-informed threat modeling. By training on annotated social media data and evaluating against traditional classifiers and deep learning baselines, the model is assessed for its precision, recall, F1-score, and false positive rate ,with special focus on real-world deployability. This study introduces an innovative and ethically designed AI framework that shifts school violence prevention from a reactive to a proactive strategy. It presents a prompt-tuned, cross-level transformer model capable of analyzing both individual posts and user behavior over time. The system uniquely integrates a domain-specific behavioral lexicon to enhance interpretability and threat detection accuracy. Moreover, it introduces new contextual variables—such as escalation patterns and term clustering—not commonly explored in prior research. Using real-world data from Reddit and YouTube, the model demonstrates how deep learning can power real-time, scalable early warning systems for public safety.