Online Opinion Trend Prediction for Public Health Events Based on Time Series Transformer

Accurate prediction of public opinion trends during public health emergencies is crucial for understanding public attitudes and enabling proactive responses. Existing methods frequently exhibit inadequate prediction accuracy and elevated computational complexity in long-term forecasting. The proposed model is an enhanced time series transformer model that incorporates three key innovations. First, a sparse probabilistic attention mechanism reducing spatial complexity from $\mathbf{O(L^{2})}$ to $\mathbf{O(LlnL)}$. Second, a progressive sequence decomposition architecture that explicitly separates trend and seasonal components. Third, a global attention distillation technique to mitigate error accumulation in autoregressive prediction. Experiments on a COVID-19 Weibo dataset containing over 780 000 posts demonstrate that the model accurately predicts trends up to seven times the input sequence length. The model outperforms existing methods by over 20% in terms of mean squared error (MSE) and mean absolute error (MAE). For a prediction length of 720, the model achieves an MSE of 0.457 and an MAE of 0.373, effectively capturing key fluctuation patterns and peak timings. The findings establish a substantial technical basis for public health management early-warning systems.