A Hybrid Model Based on Improved Transformer and Graph Convolutional Network for COVID-19 Forecasting.

Area of Science:

• Epidemiology and Public Health
• Computational Biology and Bioinformatics
• Artificial Intelligence and Machine Learning

Background:

• The COVID-19 pandemic has demonstrated significant global impact and fatality rates.
• Effective pandemic control has been challenged by the rapid spread and difficulty in accurate trend prediction.
• Existing traditional and deep learning models exhibit limitations in COVID-19 forecasting accuracy.

Purpose of the Study:

• To develop an advanced hybrid model for accurate COVID-19 trend forecasting.
• To address the limitations of low prediction accuracy in current forecasting models.
• To provide a more precise benchmark for epidemic control strategies.

Main Methods:

• Proposed a hybrid model integrating an improved Transformer with a Graph Convolutional Network (GCN).
• Utilized a multi-head attention mechanism for rich temporal sequence information extraction.
• Employed cosine function instead of softmax to reduce Transformer's time complexity from O(N^2) to O(N).

Main Results:

• The hybrid model demonstrated superior predictive performance compared to existing deep learning and traditional models.
• The model achieved faster convergence than current deep learning models.
• Evaluated using Mean Absolute Percentage Error and Mean Absolute Error on three US states with varying impact levels.

Conclusions:

• The proposed hybrid Transformer-GCN model offers enhanced accuracy and efficiency for COVID-19 forecasting.
• This improved forecasting capability serves as a valuable tool for public health decision-making and epidemic control.
• The model's reduced time complexity makes it a more practical solution for real-time pandemic monitoring.