Digital Persuasion: Understanding the Impact of Online Influencers on Public Opinion
Omran Berjawi, Rida Khatoun, Giuseppe Fenza
cs.SI · cs.LG
Quick verdict
Ranks influencers via FJ fundamental matrix and simulates opinion shifts on 2020 US-election tweets; promising, but significance untested, baselines thin.
Summary
The paper proposes a framework built on the Friedkin-Johnsen (FJ) opinion dynamics model to identify influential users and quantify their effect on community-level opinion. Using a public Kaggle dataset of roughly 1.7 million tweets from the 2020 U.S. presidential election, the authors build a directed mention graph (edge weights equal mention frequency), assign each user an initial opinion derived from RoBERTa sentiment scores, and compute equilibrium opinions as x=(I+L)^-1*s. Influencers are ranked by an InScore, the row average of the fundamental matrix, and the top 15 are split into three five-user batches. In intervention experiments, initial opinions of these users or a random batch are set to +1 or -1 and equilibria recomputed: Batch 1 moves the network mean from -0.3932 to 0.0866 (positive) and to -0.5857 (negative); Batches 2 and 3 shift less; the random batch is nearly inert (-0.3904/-0.3940). Propagation analysis finds the top-10 most affected users include indirectly connected accounts for Batch 1, but mostly direct neighbors for Batch 2. Key limitations: “significant” effects claimed without statistical tests or repeated random draws; Louvain communities detected but never used; no graph statistics or released code; only one centrality baseline; sentiment polarity conflated with political stance.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An InScore defined as the row average of the FJ fundamental matrix identifies and ranks the same top influencers as normalized centrality while providing a graded score (§4.2, Table 1).
- Manipulating the initial opinions of the top-5 influencers moves the whole-network equilibrium mean from -0.3932 to 0.0866 (positive) or -0.5857 (negative), versus a near-inert random batch (-0.3904/-0.3940) (§4.3, Table 2).
- Influence of the highest-scoring influencers extends beyond direct neighbors to neighbors-of-neighbors, whereas ranks 6-10 mostly affect direct neighbors (§4.4, Table 3).
- InScore correlates positively with follower count and total retweets but not with tweet volume (§4.2, Table 1).
Strengths
Structurally principled influence score
InScore (Eq. 3) is derived from the FJ fundamental matrix Ω=(I+L)^-1*s at equilibrium, capturing global network sensitivity rather than purely local heuristics; Table 1 shows the FJ ranking coincides with the normalized-centrality ranking of ref [2] while adding a graded score, and correlates qualitatively with follower count and total retweets.
Clean counterfactual intervention design
Setting initial opinions s_i=±1 for ranked influencer batches vs a random batch (§3.3) produces a monotone dose-response in Table 2: Batch 1 shifts the equilibrium mean by +0.48/-0.19 from -0.3932, Batches 2-3 progressively less, and the random batch by ≤0.003, supporting the framework's internal logic.
Propagation analysis beyond immediate neighbors
Public, reusable data source
Weaknesses
No statistical inference despite “significant” claims
Possible directionality error in InScore
Unused pipeline component and missing reporting basics
Thin baseline comparison and construct validity gap
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (cs.SI/cs.LG) proposes a framework built on the Friedkin-Johnsen (FJ) opinion dynamics model with three declared contributions (§1): influencer identification via the fundamental matrix at equilibrium, sentiment-manipulation experiments contrasting influential vs random users, and propagation analysis of who is most affected. The empirical setting is a public Kaggle dataset of >1.7M tweets spanning roughly one month of the 2020 U.S. presidential election (§4.1, ref [10]). The central claims are that (i) an InScore defined from the FJ fundamental matrix ranks influencers consistently with normalized centrality (§4.2, Table 1); (ii) manipulating the top-5 influencers' initial opinions shifts the network's equilibrium mean far more than manipulating random users (§4.3, Table 2); and (iii) high-InScore influence reaches neighbors-of-neighbors whereas mid-ranked influence stays local (§4.4, Table 3).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper actually do?
It builds a simulated pipeline on a public dataset of about 1.7 million tweets from the 2020 U.S. election: users are connected by mention frequency, each user's sentiment (from a RoBERTa model) becomes an initial opinion, and the Friedkin-Johnsen opinion dynamics model computes a network-wide equilibrium. Users are ranked by an influence score derived from that equilibrium, and the paper simulates what happens to the network's average opinion if selected users are forced to extreme positive (+1) or negative (-1) positions.
How much did top influencers change overall opinion in this study?
In simulation, setting the initial opinions of the top-5 ranked influencers to +1 shifted the network's average equilibrium opinion from -0.3932 to 0.0866; setting them to -1 shifted it to -0.5857. The same intervention on randomly chosen users moved the mean only to -0.3904 or -0.3940 — a change under 0.003. These are simulation outcomes, not measurements of real persuasion.
Did influencer impact spread beyond direct neighbors?
The paper's propagation analysis (Table 3) reports that for the top-5 influencer batch, 3-4 of the 10 most-affected users were indirectly connected (neighbors of neighbors), while for influencers ranked 6-10, 8-9 of the 10 most-affected users were direct neighbors. The counts come from top-10 lists and are not statistically tested.
Are the data and code available?
The tweet dataset is public on Kaggle (“US Election 2020 Tweets”, ~1.7M tweets). The paper does not mention releasing its own code, and key details such as final graph size and random-user selection are not reported, so exact reproduction would require reimplementation from the method description.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-2 (opinion distributions before/after manipulation) are referenced but cannot be assessed from extracted text; the review relies on the numeric tables only.
- No supplementary material appears in the extraction, so implementation details (graph size, random-batch sampling) may exist that could not be reviewed.
- The running heads contain a template placeholder (“Title Suppressed Due to Excessive Length”), suggesting the source PDF was an unfinished build; content otherwise appears complete (all sections, tables, and references present).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.