Analysis and Early Detection of Rumors in a Post Disaster Scenario

Abstract

The use of online social media for post-disaster situation analysis has recently become popular. However, utilizing information posted on social media has some potential hazards, one of which is rumor. For instance, on Twitter, thousands of verified and non-verified users post tweets to convey information, and not all information posted on Twitter is genuine. Some of them contain fraudulent and unverified information about different facts/incidents - such information are termed as rumors. Identification of such rumor tweets at early stage in the aftermath of a disaster is the main focus of the current work. To this end, a probabilistic model is adopted by combining prominent features of rumor propagation. Each feature has been coded individually in order to extract tweets that have at least one rumor propagation feature. In addition, content-based analysis has been performed to ensure the contribution of the extracted tweets in terms of probability of being a rumor. The proposed model has been tested over a large set of tweets posted during the 2015 Chennai Floods. The proposed model and other four popular baseline rumor detection techniques have been compared with human annotated real rumor data, to check the efficiency of the models in terms of (i) detection of belief rumors and (ii) accuracy at early stage. It has been observed that around 70% of the total endorsed belief rumors have been detected by proposed model, which is superior to other techniques. Finally, in terms of accuracy, the proposed technique also achieved 0.9904 for the considered disaster scenario, which is better than the other methods.