Concept Drift in Japanese COVID-19 Infection Data

Abstract

In this study, we analyze concept drifts in the daily infection data of COVID-19 in Japan. A lockdown, the spread of vaccines, and the emergence of new variants of COVID-19 have had a significant impact on the number of daily infections. These changes, also known as concept drifts, make the prediction of COVID-19 infection rates difficult. Because the prediction of infection trends is crucial to protect people from the disease, this study aims to generate accurate predictions by handling concept drifts in the trend data. The key concept behind this method is a brute-force tuning of the training period. Although prior studies tended to require pre-tuned parameters to locate the drift points, this can be avoided through brute-force tuning. Experimental results show significant improvements in prediction accuracy. Furthermore, the extracted points where concept drifts occur appear to correspond to new COVID-19 variants and other important state changes.