Audience Analysis for Competing Memes in Social Media

Samuel Carton, Souneil Park, Nicole Zeffer, Eytan Adar, Qiaozhu Mei, Paul Resnick

Abstract

Existing tools for exploratory analysis of information diffusion in social media focus on the message senders who actively diffuse the meme. We develop a tool for audience analysis, focusing on the people who are passively exposed to the messages, with a special emphasis on competing memes such as propagations and corrections of a rumor. In such competing meme diffusions, important questions include which meme reached a bigger total audience, the overlap in audiences of the two, and whether exposure to one meme inhibited propagation of the other. We track audience members’ states of interaction, such as having been exposed to one meme or another or both. We analyze the marginal impact of each message in terms of the number of people who transition between states as a result of that message. These marginal impacts can be computed efficiently, even for diffusions involving thousands of senders and millions of receivers. The marginal impacts provide the raw material for an interactive tool, RumorLens, that includes a Sankey diagram and a network diagram. We validate the utility of the tool through a case study of nine rumor diffusions. We validate the usability of the tool through a user study, showing that nonexperts are able to use it to answer audience analysis questions.

Introduction

Most social media analytics tools focus on propagator analysis: identifying who is tweeting, how often, and who is retweeting them. While useful, this perspective ignores the vast majority of users who see a message but never interact with it. This "sender-focused" approach leaves critical questions unanswered regarding the true reach of a rumor and the effectiveness of debunking efforts. Tracking the passive audience is computationally challenging because it requires processing massive follower lists and accounting for overlapping audiences, where one user follows multiple people spreading the same meme.

The RumorLens System

To address these challenges, we developed RumorLens, a tool that shifts the focus from senders to the audience. The system defines seven possible states for a user, such as "No exposure," "Exposed to Rumor," "Exposed to Both," and "Tweeted Rumor." The core technical innovation is the calculation of marginal impact: instead of just counting followers, the system calculates exactly how many new people a specific tweet moved from one state to another.

The RumorLens interface featuring the Sankey diagram (A, B) showing user transitions between exposure states.

RumorLens consists of a precomputation stage using a one-pass algorithm ($O(nd_{AVG})$ complexity) and an interactive visualization stage. The interface features a Sankey diagram to show the flow of users between states, a network diagram, and a raw data table. This allows researchers to observe the "inoculation" effect—where a correction reaches a user before the rumor—and track the lag between rumor propagation and debunking efforts.

Key Insights: Reach and Overlap

We evaluated RumorLens using nine real-world rumor diffusions, including the "Jay-Z is dead" hoax and various Boston Marathon bombing rumors. A primary finding is that naively summing follower counts leads to massive overestimation of reach. Because traditional methods fail to account for audience overlap, they overestimate true reach by anywhere from 6% to nearly 400%.

Comparison of marginal exposure versus raw follower counts across nine rumor diffusions, highlighting the significant overestimation inherent in naive follower counts.

Furthermore, the data suggests that opinions are often "sticky." In the Jay-Z case study, only a tiny fraction of users who tweeted the rumor later tweeted the correction, even if they were exposed to both. Reach disparity also varied significantly; for instance, the "Sandy Hook Principal" rumor had 30 rumor exposures for every 1 correction exposure, while the "HIV in Greece" rumor had a much closer 1.5 to 1 ratio, likely due to active participation by the World Health Organization in the correction.

Conclusion

RumorLens demonstrates that audience-centric analysis provides a more accurate picture of a meme's true reach and the effectiveness of debunking efforts. In a user study with 31 non-experts, the Sankey diagram was significantly more effective for answering audience-related questions than traditional diagrams, yielding an 85% mean accuracy. While the current model uses follower counts as a proxy for exposure, it provides a robust framework for understanding the dynamics of competing memes in social media.

Summary updated

Abstract

Existing tools for exploratory analysis of information diffusion in social media focus on the message senders who actively diffuse the meme. We develop a tool for audience analysis, focusing on the people who are passively exposed to the messages, with a special emphasis on competing memes such as propagations and corrections of a rumor. In such competing meme diffusions, important questions include which meme reached a bigger total audience, the overlap in audiences of the two, and whether exposure to one meme inhibited propagation of the other. We track audience members’ states of interaction, such as having been exposed to one meme or another or both. We analyze the marginal impact of each message in terms of the number of people who transition between states as a result of that message. These marginal impacts can be computed efficiently, even for diffusions involving thousands of senders and millions of receivers. The marginal impacts provide the raw material for an interactive tool, RumorLens, that includes a Sankey diagram and a network diagram. We validate the utility of the tool through a case study of nine rumor diffusions. We validate the usability of the tool through a user study, showing that nonexperts are able to use it to answer audience analysis questions.
BibTeX
@inproceedings{carton2021audience,
  author = {Samuel Carton and Souneil Park and Nicole Zeffer and Eytan Adar and Qiaozhu Mei and Paul Resnick},
  title = {{Audience Analysis for Competing Memes in Social Media}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/icwsm.v9i1.14632},
}
Figure