Using Fuzzy Clustering with Deep Learning Models for Detection of COVID-19 Disinformation

Abstract

Since the beginning of 2020, the COVID-19 pandemic has killed millions of people around the world, leading to a worldwide panic that has fueled the rapid and widespread dissemination of COVID-19-related disinformation on social media. The phenomenon, described by the World Health Organization (WHO) as an "indodemic" presents a serious challenge to governments and public health authorities, but the spread of misinformation has made human detection less efficient than the rate of spread. While there have been many studies developing automated detection techniques for COVID-19 fake news, the results often refer to high accuracy but rarely to model detection time. This research uses fuzzy theory to extract features and uses multiple deep learning model frameworks to detect Chinese and English COVID-19 misinformation. With the reduction of text features, the detection time of the model is significantly reduced, and the model accuracy does not drop excessively. This study designs two different feature extraction methods based on fuzzy classification and compares the results with different deep learning models. BiLSTM was found to provide the best detection results for COVID-19 misinformation by directly using deep learning models, with 99% accuracy in English and 86% accuracy in Chinese. Applying fuzzy clustering to English COVID-19 fake news data features maintains 99% accuracy while reducing detection time by 10%. For Chinese misinformation, detection time is reduced by 15% at the cost of an 8% drop in accuracy.