Using impression data to improve models of online social influence

Rui Liu, Kevin Greene, Mihovil Mandic, Benjamin Valentino, Soroush Vosoughi, V. S. SubrahmanianScientific 2021 · DOI

Abstract

Influence, the ability to change the beliefs and behaviors of others, is the main currency on social media. Extant studies of influence on social media, however, are limited by publicly available data that record expressions (active engagement of users with content, such as likes and comments), but neglect impressions (exposure to content, such as views) and lack "ground truth" measures of influence. To overcome these limitations, we implemented a social media simulation using an original, web-based micro-blogging platform. We propose three influence models, leveraging expressions and impressions to create a more complete picture of social influence. We demonstrate that impressions are much more important drivers of influence than expressions, and our models accurately identify the most influential accounts in our simulation. Impressions data also allow us to better understand important social media dynamics, including the emergence of small numbers of influential accounts and the formation of opinion echo chambers.

Introduction

Measuring social influence is notoriously difficult because researchers usually only have access to public-facing data. On platforms like Twitter or Facebook, we can see who liked a post, but we cannot see the thousands of people who read it and moved on. This "expression-only" view ignores the vast majority of user activity. Furthermore, observational studies lack "ground truth." If a user changes their opinion, it is nearly impossible to prove which specific post or account caused that shift using public data alone. This study addresses these gaps by creating a controlled environment where every view (impression) and every interaction (expression) is recorded, alongside daily surveys to track actual changes in user opinion.

Methodology: The DartPost Platform

The researchers developed a custom, web-based micro-blogging platform called DartPost, designed to mimic the functionality of Twitter. Using this platform, they recruited 287 participants via Amazon Mechanical Turk, assigning 200 as "influencers" and 87 as "observers." Over five days, users discussed controversial topics while observers reported their stances and nominated influential accounts in daily surveys.

To analyze this data, the authors implemented several novel modeling concepts:

  • Direct Influence Networks (DINs): These trace an "influence event" (when a user nominates someone as influential) back to the specific posts that user was exposed to within a specific time window.
  • Full Influence Networks (FINs): An expansion of DINs that recursively traces the history of posts—for example, including the original post that a viewed comment was responding to—to create a complete map of information flow.
Direct Influence Network (DIN) structure showing the relationship between posts (expressions) and views (impressions) over time.

To validate these models, the authors used Quasi-Poisson and Gamma-Poisson regression models to determine which factors (posts, likes, or views) actually predicted a user being named "influential."

Key Insights: The "Iceberg" of Social Media

The study reveals that impressions are far more frequent than expressions, representing the submerged portion of the social media "iceberg." In the simulation, there were 35.6 times more nodes in influence networks when impressions were included compared to expression-only models.

Comparison of network models: (a) Original DIN, (b) DIN incorporating impression data, and (c) the Full Influence Network (FIN).

When all variables were analyzed together, only "Received Views" (impressions) remained a statistically significant predictor of influence. Traditional metrics like "likes" or "comments" became insignificant by comparison. Regression analysis showed that increasing an account's impressions by just one standard deviation more than doubled the expected number of times users would cite that account as influential.

Results and Discussion

The proposed influence models (specifically FI1, which is network-centric, and FI3, which combines network importance with temporal decay) achieved over 77% precision at K=10. This indicates they were highly accurate at identifying the top 10 most influential accounts as identified by the users themselves.

Beyond individual influence, the data revealed rapid polarization. Echo chambers formed almost immediately; as the simulation progressed, the diversity of content users viewed dropped sharply. Users quickly settled into feeds that reinforced their existing views, filtering out opposing polarities. Additionally, influence proved to be "sticky." Accounts that were not deemed influential on the first day rarely gained influence later, with 65% of Day 1 "un-influential" accounts remaining so for the entire study.

Conclusion

This research demonstrates that impressions—simply viewing content—are the primary drivers of online influence. By incorporating view data, models can more accurately identify influential users and explain the formation of polarized echo chambers. Future work could utilize the DartPost framework to study the impact of social media bots or the success of specific disinformation strategies in environments where impression data is fully transparent to researchers.

Summary updated

Abstract

Influence, the ability to change the beliefs and behaviors of others, is the main currency on social media. Extant studies of influence on social media, however, are limited by publicly available data that record expressions (active engagement of users with content, such as likes and comments), but neglect impressions (exposure to content, such as views) and lack "ground truth" measures of influence. To overcome these limitations, we implemented a social media simulation using an original, web-based micro-blogging platform. We propose three influence models, leveraging expressions and impressions to create a more complete picture of social influence. We demonstrate that impressions are much more important drivers of influence than expressions, and our models accurately identify the most influential accounts in our simulation. Impressions data also allow us to better understand important social media dynamics, including the emergence of small numbers of influential accounts and the formation of opinion echo chambers.
BibTeX
@inproceedings{liu2021using,
  author = {Rui Liu and Kevin Greene and Ruibo Liu and Mihovil Mandic and Benjamin Valentino and Soroush Vosoughi and V. S. Subrahmanian},
  title = {{Using impression data to improve models of online social influence}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-96021-3},
}
Figure