Credibility Evaluation of Twitter-Based Event Detection by a Mixing Analysis of Heterogeneous Data

Description

Twitter-based event detection systems use information in tweets posted on Twitter by Twitter users for detecting events Twitter has a simple registration and post system, which means that anyone can post tweets freely with only a simple registration process. Twitter has been recognized as an important data resource for real-time event detection. However, Twitter-based event detection systems cannot guarantee credibility in terms of their detection results. Rumor detection has been studied recently to enable credible event detection. More specifically, the existing studies detect rumors by identifying and checking special features of incredible information on Twitter. However, values of the identified features can be faked easily, so that it is important to conduct a mixing analysis of both Twitter and external credible data resources to solve this problem. To solve this problem, this paper proposes a method to evaluate the credibility of Twitter-based event detection, which considers the two kinds of data resources for credibility evaluation to exclude influence by falsification. The experiments show that the proposed method gives correctly detected events high credibility and the other events low credibility. More specifically, event detection accuracy increases by an average of 26.8 % by reviewing the detection results according to their credibility evaluated with the proposed method.