Deep-learning based
Covid-19 forecasting
with model-based data augmentation
Fong, Chung Yan, and Dit-Yan Yeung
"Deep COVID-19 Forecasting for Multiple States with Data Augmentation"
arXiv:2302.01155 (2023)
Key ideas
Transformer model
The self-attention mechanism of the transformer allow multiple time series to interact with each other.
Ensemble method
By combine models which are randomly initialized into an ensemble, the overall forecast become more robust.
Data augmentation
Employing an SIR-based data model, we generate more data for training and validation. The generated data was shown to be distributionally closer to the test data.
