Time Warping Clustering for the Forecast and Analysis of COVID-19

Abstract

This paper presents an effective algorithm for the clustering of confirmed COVID-19 cases at the county-level in the United States. Dynamic time warping and Euclidean distance are examined as the k-means clustering distance metrics. Dynamic time warping can compare time series varying in speed, as counties often experience similar outbreak trends without the timelines matching up exactly. The effect of data preprocessing on clustering was systematically studied. Further analyses demonstrate the immediate value of our clusters for both retrospective interpretation of the pandemic and as informative inputs for case prediction models. We visualize the time progression of COVID-19 from April 5, 2020 to August 23, 2020. We proposed a Monte-Carlo dropout feedforward neural network with the ability to forecast four weeks into the future. Predictions evaluated from July 24, 2020 to August 20, 2020 demonstrate the better empirical performance of the model when trained on the clusters, in comparison with the model trained on individual counties and the model trained on counties clustered by state.