An Unsupervised Misinformed Users and Content Identification of Covid-19 on Twitter Leveraging Frequency of Posts

Abstract:

In March, 2020, Covid-19 was declared an infodemic by the experts in the 86th report situational report published by WHO, "Infodemic is defined as an overabundance of information-some accurate and some not, that makes it hard to find trustworthy sources and reliable guidance." A study by the Reuters Institute for the Study of Journalism found that 40 % believed that social media is the primary source of the spread of misinformation. Some research works have focused on analyzing the misinformed Twitter data using bot detection and fake-URL identification tools. However, these methods require annotated data in advance and high capacity hardware resources to deal with vast amounts of data. We propose an unsupervised framework to detect misinformed content and the users who may be the sources of or susceptible to spreading misinformed content using approx. 51 million tweets collected from 27 March to 10 April 2020. The proposed framework extracts the influential clusters consisting of the tweets retweeted most frequently over a given duration. First the spherical word embeddings of the tweets are generated and then clustered using SK-means clustering. The combination of spherical embeddings and SK-means show better performance on the text clustering tasks because they capture the directional information of vectors and the contextual information from the text. The influential clusters are further filtered using the user-level information from four categories of users i.e., News, Medical Organization, Medical Influencers, and Verified users. These categories are dynamically updated by using the semantic similar textual information from the user-descriptions of these user categories. The framework has been tested on two public annotated datasets from FakeNewsNet, which gives an average 13% improvement in F1-score from Multinomial Naive Bayes, Gaussian Naive Bayes and SVM using TF-IDF and Word2Vec. Due to the unavailability of the annotated Twitter data, the resulting suspected misinformed users were searched manually on Twitter, where approximately 50% identified users were found already suspended. The content of misinformed clusters was validated by performing emotional analysis of the tweets. This showed that the suspected misinformed clusters contained 70% anger, 15 % fear, 7% sad and 8% joy presenting tweets.