A Dual-Channel Transformer-SHAP Framework for Early Detection and Interpretable Analysis of Youth Social Burnout from Multimodal Behavioral Data

Abstract

An efficient and interpretable warning model is urgently needed to address the prominent risk of social burnout among adolescents in high-intensity social media interactions. Traditional machine learning methods have insufficient modeling capabilities for temporal behavior features and lack interpretability in prediction. This study proposes a warning method that integrates Transformer and SHAP: a dual channel Transformer architecture is designed, where the main channel analyzes the high-order correlation of 15-dimensional user behavior sequences through stacked encoding layers and multi head attention, and the auxiliary channel uses a recursive feature pyramid structure to enhance state sensitivity; Furthermore, fatigue risk prediction can be achieved through a fully connected network. Simultaneously constructing a tree structure optimized SHAP calculation process, while reducing high-dimensional computational complexity, verifying the rationality of feature attribution through psychological theory, and using SHAP to preprocess abnormal samples, combined with variational autoencoder (VAE) to achieve deviation detection, forming an integrated mechanism of "warning traceability". Based on a desensitization set containing 6-month social data of 5892 adolescents, the robustness was verified through missing data and data change scenarios. The results showed that the model had an accuracy rate of 92.37% (improved by 8.5% and 12.2% compared to LSTM and random forest, respectively), an F1 value of 0.891, an early fatigue recall rate of 85.2%, a false positive rate of<7.3%, and a response time of 1.8 seconds; SHAP analysis confirmed that nighttime activity exceeding 2 hours (+0.32), weekly social decline exceeding 40% (+0.28), and negative emotional words exceeding 15% (+0.41) are the core features, with a combined contribution of 68.7%, providing quantitative evidence for targeted interventions.