Topic Detection in COVID-19 Mortality Time Series

The mortality of COVID-19 has been analyzed from a predictive point of view, building models that failed to predict the medium and long term evolution of the time series. Looked in a retrospective way, it is possible to appreciate that the mortality impact of the pandemic was very different across countries. In this paper we deal with the discovery of representative patterns, i.e. topics in latent Dirichlet analysis (LDA), that may explain the observed mortality time series. The choice of the number of topics is a balance between the minimization of perplexity, and the avoidance of overfitting. We find that countries can be clustered according to the coefficients of the decomposition of their time series into topics, and that this decomposition has a geopolitical correspondence, i.e. countries with the same dominant topic representation belong to the same geographical and/or political domain.