CoviFake: A Framework to Detect and Analyze Fake COVID19 Tweets

Along with the unprecedented impact of the COVID-19 pandemic on human lives, a new crisis of fake and false information related to disease has also emerged. Primarily, social media platforms such as Twitter are used to disseminate fake information due to ease of access and their large audience. However, automatic detection and classification of fake tweets is challenging task due to the complexity and lack of contextual features of short text. This paper proposes a novel CoviFake framework to classify and analyze fake tweets related to COVID-19 using vocabulary and non-vocabulary features. For this purpose, first, we combine and enhance ‘CTF’ and ‘COVID19 Rumor’ datasets to build our COVID19-sham dataset containing 25,388 labelled tweets. Next, we extract the vocabulary and 12 non-vocabulary features to compare the performance of six state-of-the-art machine learning classifiers. Our results highlight that the Random Forest (RF) classifier achieves the highest accuracy of 94.53% with the combination of top 2,000 vocabulary and 12 non-vocabulary features. In addition, we developed a large-scale dataset of CoviTweets containing 7.88 million English tweets posted by 3.8 million users during two months (March-April, 2020). The analysis of CoviTweets leveraging our framework reveals that the dataset contains 1.64 million (20.87%) fake tweets. Furthermore, we perform an in-depth examination by assigning a ‘fakeness score’ to hashtags and users in CoviTweets.