Explainable AI-Driven Early Warning of Adolescents' Social-Value and Emotional Risks on Social Media Using Transformer-SHAP Fusion

Abstract

This study constructed a hierarchical early warning system for identifying adolescents' social text emotions and ideological values. It was developed because existing models lacked systematic risk classification, ignored ideological dimensions, and had uninterpretable black-box defects that hindered practical educational interventions. To address this, an 82,000-sample dataset was built using a four-level risk annotation covering psychological emotions and ideological values. A lightweight transformer encoder was integrated with the Shapley-additive-explanations algorithm to quantify textual feature contributions. The proposed model achieved an accuracy of 89.42% and a macro-F1 score of 87.65%, outperforming baseline models while reducing parameters by 62% and increasing inference speed by 47% relative to bidirectional-encoder-representations-from-transformers base. Ultimately, this provided accurate, real-time, and interpretable technical support for youth mental health protection, online ideological education, and precise campus crisis interventions.