Feature Enriched Framework for Rumor Detection Using Tweets

Online social network platforms are elaborately leveraged for the purpose of information or news collection during impactful events. Such information lacks credibility and proper verification which unravel rumors and can prove to be hazardous during high-impact events. The lack of standardization of the process of restraint at the message level for online social network like Twitter further proliferates the spread of rumor among huge number of users within a short span of time. In order to minimize the negative impact of rumor spread in Twitter during impactful events, this research has proposed a feature enriched rumor detection framework capable of detecting and classifying rumors from tweets generated during high-impact events on Twitter. The features categories are: content, user and contextual. Each of the content and user features are enriched by two new features. The proposed framework also provides an integral platform where for every incoming tweet, content features, user features and contextual features are modeled by machine learning, pre-trained word embedding with deep learning approaches and transformer models. The experiments were performed in two ways on benchmark rumor datasets Pheme and Twitter16. The experimental results show XGBoost, ensemble technique, word embedding with BiLSTM and XLNet models that are the most effective in detecting and classification of rumors. The enrichment features helped to achieve a notable performance in accuracy and recall measures of 89% and 91%, respectively.