Forecasting the COVID-19 Space-Time Dynamics in Brazil With Convolutional Graph Neural Networks and Transport Modals

Abstract

One of the major challenges imposed by the SARS-CoV-2 pandemic is the lack of pattern in which the virus spreads, making it difficult to create effective policies to prevent and tackle the pandemic. Several approaches have been proposed to understand the virus behavior and anticipate its infection and death curves at country and state levels, thus supporting containment measures. Those initiatives generalize well for general extents and decisions, but they do not predict so well the trajectory of the virus through specific regions, such as municipalities, considering their distinct interconnection profiles. Especially in countries with continental dimensions, like Brazil, too general decisions imply that containment measures are applied either too soon or too late. This study presents a novel scalable alternative to forecasting the numbers of cases and deaths by SARS-CoV-2, according to the influence that certain regions exert on others. By exploiting a single-model architecture of graph convolutional networks with recurrent networks, our approach maps the main access routes to municipalities in Brazil using the modals of transport and processes this information via neural network algorithms to forecast at the municipal level and for the whole country. We compared the performance in forecasting the pandemic daily numbers with three baseline models using Mean Absolute Error (MAE), Symmetric Mean Absolute Percentage Error (sMAPE) and Normalized Root Mean Square Error (NRMSE) metrics, with the forecasting horizon varying from 1 to 25 days. Results show that the proposed model overcomes the baselines when considering the MAE and NRMSE ( $p$ - $value < 0.01$ ), being especially suitable for forecasts from 14 to 24 days ahead.

Keywords

• Covid-19
• graph convolutional networks
• time series forecasting
• transport modals