Fake News Detection of South African COVID-19 Related Tweets using Machine Learning

Social Media has grown in popularity in recent years comprising of billions of users who in turn exchange and communicate content at a volume and rate impractical to examine manually. Fake News are now being used on these platforms to manipulate and affect societies across the world as was the case in the 2016 United States of America (USA) elections and of recent during the 2019 coronavirus disease (COVID-19) pandemic. South Africa is not immune to the spread of Fake News, particularly, through Social Media platforms such as Twitter, Facebook and TikTok. It is, therefore, important to detect the presence of Fake News computationally in order assist the mitigation of its spread and prevent perceivable negative effects. This study addresses the issue by developing a Machine Learning (ML) model to analyze large amounts of data associated with Social Media. Curated annotated datasets from CONSTRAINT AAAI 2021; COVID-19 Rumour, FNIR and Zenodo’s COVID-19 datasets; Google and Polifact Fact Checked websites; were utilized to develop the ML model. Specifically, the model was trained on 36254 data points and applied on a South African related COVID-19 Twitter dataset collected for cursory analysis. In total, 27 ML models were experimented with and the collected South African COVID-19 related Twitter dataset comprised of 976087 tweets from 8 November 2020 until 19 July 2021. The results detected 329107 tweets as being ‘Fake’ based on the LightGBM Classifier which was chosen as the most feasible model in terms of speed and a balanced accuracy score of 0.82. The proposed model is unique as it is trained on a larger combined dataset and supplements existing efforts to combat misinformation, disinformation and malinformation spread on Twitter.