Deep Learning-Based Prediction of Polarization in Social Media Among Youth

The stochastic nature of youth social media polarization poses challenges for traditional early-stage prediction. This study constructs a heterogeneous temporal network to represent multifaceted interactions among users, text, and hashtags, treating the evolution of online opinion as a spatiotemporal signal-propagation process in networked communication environments. Based on the DyGCN-GRU framework, the model captures the spatiotemporal propagation patterns of public-opinion evolution, integrates sentiment kurtosis, semantic variance, and community opposition density to generate polarization-intensity sequences, and finally employs a Transformer decoder to predict polarization risk. The network incorporates platform-specific embeddings, youth-oriented semantic features, and behavioral attributes to improve the detection of weak early signals. Experimental results on Weibo and Zhihu data show that the proposed method achieves an F1-score of 0.915, a root mean square error of 0.068, and an early-warning lead time of 11.3 minutes. The results verify the model’s ability to quantitatively perceive and dynamically predict early risks of youth public-opinion polarization, and provide a technical basis for online risk monitoring and intelligent information-governance systems.