Study on the rumor detection of social media in disaster based on multi-feature fusion method

Abstract

Abstract In recent years, there is a significant increase in research combining social media data for disaster warning and damage assessment. When natural disasters occur, social media data can also contain rumors, which not only reduce the accuracy of assessment but also have a very negative social impact. In this paper, a multi-feature fusion neural network with attention mechanism is proposed for rumor detection, which makes the attempt to integrate user, textual and propagation features in one united framework. Specifically, a Bi-directional Long Short Term Memory Network (Bi-LSTM) is applied to extract user and textual features and a Graph Convolutional Neural Network (GCN) is employed to extract the high-order propagation features. In addition, both the complementary and alignment relationships between different features are considered to achieve better fusion. It shows that our method can detect rumors effectively and perform better than previous methods on the Weibo dataset. To validate the effectiveness of our model, rumor detection is conducted in the social media data collected from Typhoon Lekima on Aug 10th- 14th 2019 in China, the earthquake of magnitude 6.8 on Sep 5th- 9th, 2022 in Sichuan, China, the wildfire on Aug 15th- 26th, 2022 in Chongqing, China. Results show that: 1) the proposed method performs well in rumor detection in disaster; 2) rumors often appear along with hot topics; 3) rumors express much negative sentiment; 4) rumor propagation networks have tighter structure and deeper propagation depth. 5) rumors account for a relatively small percentage of social media data in disaster, which means that most social media data is credible.