A Graph SIR Network Based on Dynamic Graph Structures and Residual Learning for Epidemic Prediction
Lingfeng Miao, Yufan Chen, Jiawei Wang, Choujun Zhan, Xuejiao Zhao · IEEE Transactions on Consumer Electronics · 2025
According to the World Health Organization (WHO), COVID-19 has resulted in approximately 7 million deaths worldwide, posing a severe threat to public health. Accurately predicting COVID-19 infection trends can assist governments in developing strategies to mitigate the impact. This paper introduces a novel hybrid machine learning model, RLG-SIR-Net, proposed for predicting daily confirmed COVID-19 cases. DLinear is used to decompose time series data, obtaining a trend sequence and a residual sequence. The dynamic graph learning module can construct a dynamic graph from the trend sequence. Then, a graph convolutional network is adopted to extract correction information from the dynamic graph and the residual sequence. Finally, the correction information is employed to enhance the predictive performance of the SIR model. COVID-19 datasets containing data on four countries and five baseline models were used to validate the predictive performance of RLG-SIR-Net. Experimental results show that RLG-SIR-Net outperforms the other baseline models in long-term forecasting of COVID-19 infections.