Propagation Dynamics of Rumor vs. Non-rumor across Multiple Social Media Platforms Driven by User Characteristics

TL;DR

It is found that reputable active users, termed `onlookers', inadvertently or unwittingly spread rumors due to their extensive online interactions and the allure of sensational fake news, while celebrities exhibit caution, mindful of releasing unverified information.

Abstract

Studying information propagation dynamics in social media can elucidate user behaviors and patterns. However, previous research often focuses on single platforms and fails to differentiate between the nuanced roles of source users and other participants in cascades. To address these limitations, we analyze propagation cascades on Twitter and Weibo combined with a crawled dataset of nearly one million users with authentic attributes. Our preliminary findings from multiple platforms robustly indicate that rumors tend to spread more deeply, while non-rumors distribute more broadly. Interestingly, we discover that the spread of rumors is slower, persists longer, and, in most cases, involves fewer participants than that of non-rumors. And an undiscovered highlight is that reputable active users, termed `onlookers', inadvertently or unwittingly spread rumors due to their extensive online interactions and the allure of sensational fake news. Conversely, celebrities exhibit caution, mindful of releasing unverified information. Additionally, we identify cascade features aligning with exponential patterns, highlight the Credibility Erosion Effect (CEE) phenomenon in the propagation process, and discover the different contents and policies between the two platforms. Our findings enhance current understanding and provide a valuable statistical analysis for future research.

Figures (8)

• Figure 1: Complementary cumulative distribution functions (CCDFs) statistics of critical network topology-based attributes in rumor and non-rumor cascades for Twitter (the first row) and Weibo (the second row). From left to right, the plots represent (a) the maximum breadth, (b) the structural virality score, (c) the diameter, and (d) the cascade size. In all cases, the CCDFs related to broadcasting (Max-Breadth) for non-rumors are higher than that for rumors. In contrast, rumors are beyond non-rumors when associated with viral diffusion (Structural Virality and Diameter). About 90% of rumors involve fewer participants compared to non-rumors. However, the average number of participants in rumors is larger than that in non-rumors.
• Figure 2: Comparison between broadcast and viral diffusion. While broadcast diffusion typically involves a central source that spreads information uniformly to a wide audience, viral diffusion is characterized by organic, peer-to-peer spread, and information cascades often show larger depth and diameter.
• Figure 3: CCDFs of four propagation attributes for rumor and non-rumor cascades on Twitter (the first row) and Weibo (the second row). From left to right, the plots represent (a) the maximum hop distances, and (b) the average hop distances from the source user. (c) the maximum time, and (d) the average time taken for the cascades to reach other users after being spread from the source. In all cases, the CCDFs for rumor cascades are plotted above those of non-rumor cascades, indicating deeper, slower but persisting propagation patterns for rumors.
• Figure 4: CCDFs of four user attributes for rumor and non-rumor cascades on Twitter. From left to right the plots represent (a) the number of tweets, (b) the registration date, (c) the number of fans, and (d) the number of followings. In all cases, the comparison trend between rumors and non-rumors is largely inverse when observed at the average level of all participants (the first row) and at the source user level (the second row). Only the metric of followings shows an insensitive trend between rumor and non-rumor for both the source user group and the participant group.
• Figure 5: CCDFs of four user attributes for rumor and non-rumor cascades on Weibo. Similar to Fig. \ref{['fig3']}, the comparison trend between rumors and non-rumors is largely inverse when observed at the average level of all participants (the first row) and at the source user level (the second row). And only the metric of followings shows an insensitive trend between rumor and non-rumor for both the source user group and the participant group.