Related Work and Conclusions

Statistical and Machine Learning Models

Many statistical methods including Susceptible Infectious Recovery and Auto-Regressive Integrated Moving Acerage model have been used to sumulate and forecast COVID-19 spread, yet they fall short of dealing with high-dimensional and temporal data. While they are easier to interpret comparing to Deep Learning Models due to DL’s black-box nature, deep learning models excel them in terms of performance accuracy and the ability to captue non-linear complex relationships across variables. [Clement et al., 2021]

Deep Learing Models

Previously LSTM and Bi-LSTM models have demonstarted their remarkable performance because of RNN’s ability to learn from sequential data. Later on the Variational Auto Encoder has outperfoms RNN-based model. However, desptite Deep Learning Model’s outstanding performance, there are concerns about interpretability and their inability to capture socioeconomical factors. [Lim and Zohren, 2021]

Conclusions

Our proposed attention-based TFT model has achieved state-of-the-art performance while enabling new forms of interpretability through analyzing complex spatiotermporal patterns. Our model has 4 main achievements: (1) outperforms other models in all evaluation metrics (2) exhibits robust performance dealing with non-stionary data (3) interprets termporal patterns (4) interprets spatial patterns. [Vaswani et al., 2017]
Our model has creatively incorporated spatial factors at county level, and such methodology can be easily extended to other datasets at community level (e.g population, socioeconomic factors). Future work could focus on analyzing the sensitivity of input features and adaptively optimizing the model for dynamic data.