Which cascade is more decisive in rumor detection on social media: Based on comparison between repost and reply sequences

Abstract

Rumor detection research is widely carried out to control the negative impact of rumor spreading. Many researchers conduct their research based on data from social media and prefer to employ different information cascades on social media, such as reposts or replies, to detect rumors. However, most of them view reposts or replies only as input sequences and do not distinguish the two types of sequences and their effects. Hence, this paper proposes a model called “cascade-sequence-based rumor detection” (CSRD) to explore the differences between reposts and replies in rumor detection. The CSRD combines a modified dilated convolution with a bi-directional long short-term memory (Bi-LSTM) neural network to extract the local and global features from the replies or reposts. Accordingly, two rumor datasets are collected from Twitter and Weibo. Based on the two datasets, three experiments are conducted with the proposed models: the rumor detection experiment, the attention-weight experiment, and the early detection experiment. The results reveal the differences between the repost and reply sequences in rumor detection. The differences in information density of reposts and replies affect rumor detection, where replies are denser and more decisive than reposts for obtaining a higher detection accuracy. The differences between the local and global features of the reposts and replies lead to differences in the detection accuracy, where the local features substantially impact the detection results more than the global features. Different data exposure levels of reposts and replies at different propagation stages also influence the accuracy of rumor detection.