Analysis of the socioeconomic impact due to COVID-19 using a deep clustering approach

Abstract

One of the main problems that countries are currently having is being able to measure the impact of the pandemic in other areas of society (for example, economic or social). In that sense, being able to combine variables about the behavior of COVID-19 with other variables in the environment, to build models about its impact, which help the decision-making of national authorities, is a current challenge. In this sense, this work proposes an approach that allows monitoring the socioeconomic behavior of the regions/departments of a country (in this case, Colombia) due to the effect of COVID-19. To do this, an approach is proposed in which the behavior of the infected is initially predicted, and together with other context variables (climate, economics and socials) determines the current socioeconomic situation of a region. This classification of a region, with the pattern that characterizes it, is a fundamental input for those who make decisions. Thus, this work presents an approach based on machine learning techniques to identify regions with similar socioeconomic behaviors due to COVID-19, so they should eventually have similar public policies. The proposed hybrid model initially consists of a time series prediction model of infected, to which are added several context variables (climate, socioeconomic, incidence of COVID-19 at the level of deaths, suspects, etc.) in an unsupervised learning model, to determine the socioeconomic impact in the regions. Particularly, the unsupervised model groups similar regions together, and the pattern of each group describes the socioeconomic similarities between them, to help decision-makers in the process of defining policies to be implemented in the regions. The experiments showed the ability of the hybrid model to follow the evolution of the regions after 4 weeks. The quality metrics for the predictive model were around the values of 0.35 for MAPE and 0.68 for R2, and in the case of the clustering model were around the values of 0.3 for the Silhouette index and 0.6 for the Davies–Boulding index. The hybrid model allowed determining things like some regions that initially belonged to a group with a very low incidence of positive cases and very unfavorable socioeconomic conditions, became part of groups with moderately high incidences. Our preliminary results are very satisfactory since they allow studying the evolution of the socioeconomic impact in each region/department.