Machine Learning-Based Identification of Collusive Users in Twitter Stream

Twitter has become a go-to platform for most of us to share our ideas and opinions online, this has also become a place to influence people online by sharing facts and opinions. People with a higher follower count are considered to be more credible and often use this follower count as a way of determining the price of promotion for any company/individual. But buying fake followers online today is as easy as ordering a snack online therefore it is a major challenge to identify such users as they can spread fake news, incite violence, etc. We collected a dataset for a total of 1497 users consisting of 713 genuine users and 784 collusive users but many accounts were already disabled by Twitter or the user had made their account private for which cannot be accessed by Twitter API. Therefore, the final dataset was of 1066 users with 484 genuine and 598 collusive users. We propose 65 features for the analysis of users including two novel categories of features, Keyword analysis, and Timestamp Clustering analysis. In this project, we attempt to tackle this problem using machine learning models and giving users two confidence indices Personal Confidence Index (प) and Cumulative confidence index (क). We were able to achieve an accuracy of 87 percent and were able to develop an online platform Know-Your-Twitterati to analyze a user in real-time and show relevant user features on the dashboard.