Digital Persuasion: Understanding the Impact of Online Influencers on Public Opinion

1 Introduction

• The paper addresses the potential of the Friedkin-Johnsen model for identifying influential users and simulating targeted opinion manipulation.Source: §1 · p. 2
• The study uses a publicly available Kaggle dataset of tweets from the 2020 U.S. presidential election to examine how influencer interactions can change users’ opinions and convictions.Source: §1 · p. 2
• It identifies and ranks influential users in communities by interpreting the Friedkin-Johnsen model’s influence matrix at equilibrium.Source: §1 · p. 2
• It compares changes in overall opinion after manipulating influential users’ initial opinions versus those of random users.Source: §1 · p. 2
• It analyzes whether manipulation most affects influencers’ direct neighbors or also reaches neighbors of neighbors.Source: §1 · p. 2

2 Related Works

• Prior studies combine opinion dynamics with deep learning for prediction and examine influencer behavior in relation to echo chambers.Source: §2 · p. 3
• Other Friedkin-Johnsen model research considers changing susceptibility to influence, dynamic network structures, and opinion propagation.Source: §2 · p. 3
• Equilibrium-focused studies use the Friedkin-Johnsen model to analyze conditions for opinion polarization and interpret equilibrium opinions through absorbing random walks.Source: §2 · p. 3
• Where Zhang optimizes average equilibrium opinion by modifying selected nodes’ internal opinions, this work uses the fundamental matrix to identify influencers and compare targeted with random interventions.Source: §2 · p. 3

3.1 Data Preparation

• Tweet mentions define weighted interactions, while RoBERTa sentiment scores averaged per user provide their initial opinions.Mention frequencies represent interaction strength; sentiment polarity ranges from negative to positive.Source: §3.1 · p. 4

3.2 Graph Construction and Community Detection

• The directed graph links users when one mentions another, weights edges by mention frequency, and uses Louvain to detect communities.The graph is represented as G = (V, E), with users as nodes.Source: §3.2 · p. 4

3.3 Opinion Dynamics Modeling

• The FJ model keeps each user's initial opinion fixed while iteratively updating expressed opinions from weighted neighbor interactions until convergence.The equilibrium expression uses the identity matrix and Laplacian; the fundamental matrix summarizes how network structure and initial opinions shape steady-state opinions.Source: §3.3 · pp. 4–5
• Influence scores derived from fundamental-matrix elements rank users by their impact on the network's equilibrium opinions.A higher InScore_i indicates greater impact on users' final opinions.Source: Influence Identification · p. 5, §3.3 · p. 5
• The manipulation experiments compare changing influential users' versus random users' initial opinions to +1 or −1 and measuring average equilibrium opinion.The average equilibrium opinion is used to capture overall network sentiment change.Source: Opinion Manipulation · pp. 5–6

4.1 Dataset

• The dataset contains more than 1.7 million tweets collected during the 2020 U.S. election over one month.The data includes user metadata such as followers, likes, and retweets.Source: §4.1 · p. 6

4.2 Influencer Identification

• The FJ approach identified the same top 15 influencers as normalized centrality and assigned InScore rankings to them.The top 15 were divided into three batches of five by rank.Source: §4.2 · p. 6, Table 1 · p. 7
• InScore was positively correlated with follower count and retweets, but tweet count did not directly contribute to users’ influence.The paper reports that high tweet volume alone need not indicate influence.Source: §4.2 · pp. 6–7

4.3 Sentiment Manipulation

• Batch 1 positive manipulation raised average network opinion from -0.3932 to 0.0866, while negative manipulation lowered it from -0.3932 to -0.5857; higher InScore tracked stronger shifts.The top influencers exerted stronger effects than lower-ranked influencers or random users.Source: §4.3 · pp. 7, 9, Figure 1 · p. 8, Table 2 · p. 9
• Batch 2 positive and negative manipulation shifted average network opinion from -0.3932 to -0.3408 and -0.429, respectively; Batch 3 changes were minor and random-user shifts negligible.The random-user averages shifted from -0.3932 to -0.3904 and -0.3940, respectively.Source: §4.3 · pp. 7, 9, Figure 2 · p. 8

4.4 Propagation Analysis

• Batch 1 manipulation affected direct neighbors and neighbors of neighbors, whereas Batch 2’s top affected users were mostly direct neighbors (8/10 positive; 9/10 negative).The propagation analysis examined the top 10 most affected users for Batch 1 and Batch 2.Source: §4.4 · pp. 9–10, Table 3 · p. 9

• Users with higher influence ratings significantly shift community opinions.Source: §5 · p. 10
• The FJ-based framework identifies and ranks influential users by influence score through experiments manipulating their initial opinions.Source: §5 · p. 10
• Influencers’ effects extend beyond immediate neighbors to include indirect relationships.Source: §5 · p. 10
• Future research could investigate real-time opinion manipulation and validate the approach across social network platforms.Source: §5 · p. 10