Functional Clustering of Covid-19 Countrywise Pandemic Performance

Abstract

The Covid pandemic has lasted for over a year influencing everyone's physical and emotional well-beings. Our work is aimed at exploring the capability of various types of functional data clustering methods on the complex Covid data. We collect the Covid data from the Our World in Data website, where the data source is maintained by the John Hopkins University. In our study, we introduce the clustering methods that come from both non-parametric and model-based families. K-mean alignment method combines curve alignment and k-mean clustering, where there is no parametric assumptions of distribution. On the other hand, funHDDC and funFEM model the clustering on the Gaussian mixture distribution assumptions. funHDDC uses EM-like inference for parameters; funFEM is based on the Fisher EM algorithm, which combines Fisher method and EM algorithm in order to ensure the most discriminant group-specific subspace. We purposed the sequential clustering technique on the three stages of pandemic development. Model-based methods show good clustering stability on each stage compared to the non-parametric method in terms of Adjusted Rand Index (ARI). Through the mapping technique, we can conclude the clusters are very sensitive to the countries having either the most severe Covid cases or the fewest Covid cases in three algorithms. However, for countries that do not have the above extreme conditions, their clusters are unclear. The clustering algorithm, such as funFEM, would downgrade the number of clusters from three to two and others would show large variance in ARI indicating the reduction of the clustering stability.