Echo Chambers: Emotional Contagion and Group Polarization on Facebook

Abstract

Recent findings showed that users on Facebook tend to select information that adhere to their system of beliefs and to form polarized groups - i.e., echo chambers. Such a tendency dominates information cascades and might affect public debates on social relevant issues. In this work we explore the structural evolution of communities of interest by accounting for users emotions and engagement. Focusing on the Facebook pages reporting on scientific and conspiracy content, we characterize the evolution of the size of the two communities by fitting daily resolution data with three growth models - i.e. the Gompertz model, the Logistic model, and the Log-logistic model. Although all the models appropriately describe the data structure, the Logistic one shows the best fit. Then, we explore the interplay between emotional state and engagement of users in the group dynamics. Our findings show that communities' emotional behavior is affected by the users' involvement inside the echo chamber. Indeed, to an higher involvement corresponds a more negative approach. Moreover, we observe that, on average, more active users show a faster shift towards the negativity than less active ones.

Figure 1

(a) Temporal evolution of the size of the communities S₁ (solid violet), S₂ (dotted orange), S₅ (dashed pink), C₁ (solid blue), C₂ (dotted sea green), and C₅ (dashed green). (b) Boxplots of the users’ mobility within each group. From left to right, results for C₁, S₁, C₂, S₂, C₅, and S₅.

Figure 2

Fit of the temporal evolution of the size of science (a) and conspiracy (b) communities. We fitted the data with four growth models: GM (bold orange line), LM3 (dotted violet line), LM5 (dashed-dotted blue line), and LLM (dashed purple line).

Figure 3

S₅ (a) and C₅ (b) dominant spectral components. Original series are shown in orange lines, trends in dotted violet lines, and significant signal reconstructions in dashed green lines.

Figure 4. Pre-processing procedure.

(a) Detrended S₅ (solid pink) and C₅ (solid green). (b) Standardized-by-trend S₅ (solid pink) and C₅ (solid green) residual time series. In panel b, the pre-processed series are standardized to zero mean and unit variance.

Figure 5

Mean final sentiment σ_i of all users (a), science users (b), and conspiracy users (c), as a function of the user engagement. In the insets we report the value of σ_i for those users with at least 100 comments.

Figure 6

Mean negative/positive difference δ_NP(i) of all users (a), science users (b), and conspiracy users (c), as a function of the user engagement. In the insets we report the value of δ_NP(i) for those users with at least 100 comments.

Figure 7

User’s sentiment polarization ρ_σ(i) of all users (a), science users (b), and conspiracy users (c), as a function of the user engagement. In the insets we report the value of ρ_σ(i) for those users with at least 100 comments.

Figure 8

Community negative/positive difference of comments as a function of the daily community activity for science users (a,c) and conspiracy users (b,d), for all users (a,b) and users with at least 100 comments (c,d).

Figure 9

Mean community sentiment polarization as a function of the daily community activity for science users (a,c) and conspiracy users (b,d), for all users (a,b) and users with at least 100 comments (c,d).