Affective Polarization and Dynamics of Information Spread in Online Networks

TL;DR

Problem: characterize affective polarization in online networks and its impact on information diffusion. Approach: quantify emotions and toxicity in reply interactions with transformer-based detectors, infer ideology, and analyze retweet networks in two large US Twitter datasets (COVID-19 and Roe v Wade). Key results: up-group vs in-group emotion patterns, distance-dependent affect, and distinct diffusion dynamics tied to issue salience across polarized groups. Significance: links emotional dynamics to network structure and information spread, with implications for understanding and mitigating online political polarization.

Abstract

Members of different political groups not only disagree about issues but also dislike and distrust each other. While social media can amplify this emotional divide -- called affective polarization by political scientists -- there is a lack of agreement on its strength and prevalence. We measure affective polarization on social media by quantifying the emotions and toxicity of reply interactions. We demonstrate that, as predicted by affective polarization, interactions between users with same ideology (in-group replies) tend to be positive, while interactions between opposite-ideology users (out-group replies) are characterized by negativity and toxicity. Second, we show that affective polarization generalizes beyond the in-group/out-group dichotomy and can be considered a structural property of social networks. Specifically, we show that emotions vary with network distance between users, with closer interactions eliciting positive emotions and more distant interactions leading to anger, disgust, and toxicity. Finally, we show that similar information exhibits different dynamics when spreading in emotionally polarized groups. These findings are consistent across diverse datasets spanning discussions on topics such as the COVID-19 pandemic and abortion in the US. Our research provides insights into the complex social dynamics of affective polarization in the digital age and its implications for political discourse.

Figures (18)

• Figure 1: Affective polarization in networks. People express warmer feelings when replying to those closer to them in a social network; when replying to those farther away, they express more negativity.
• Figure 2: In-group and out-group affect in the Roe_v_Wade data. Boxplots show confidence scores of emotions expressed in replies interactions in the abortion discussions between same-ideology users (in-group) and opposite-ideology users (out-group). Out-group interactions show more (a) anger, (b) disgust, and use more (c) toxic language, but also slightly less (d) joy and (e) fear, and are generally (f) shorter in length (as measured by the number of characters divided by the maximum allowed length, 280 characters). The boxes span the first to third quartiles, with whiskers extending 1.5 times the interquartile range. The horizontal line inside the box represents the median, and diamond symbol marks the mean. Differences in means were tested for statistical significance using a two-sided Mann-Whitney U Test with the Bonferroni correction: * indicates significance at p$<0.05$, ** - p$<0.01$, *** - p$<0.001$, **** - p$<0.0001$ and, ns - not-significant.
• Figure 3: Affective polarization. Difference between the mean emotion confidence scores of out-group and in-group interactions in the Roe_v_Wade and COVID-19 datasets. Error bars show standard errors.
• Figure 4: Heatmap of the embeddings of social networks. Each network was constructed by linking users who retweeted each other in online discussions about a) the 2022 overturning of Roe_v_Wade and b) the COVID-19 pandemic. The massive retweet networks were embedded in a lower-dimensional space using a graph embedding method. The heatmap of the embedding shows bright spots of densely-linked communities of users who frequently retweet one another. The retweet network of abortion discussions (a) shows two overaching polarized communities, while the network of the pandemic discussions (b) has a multi-focal structure.
• Figure 5: Affective polarization in the retweet network of Roe_v_Wade data. Each plot shows emotions in reply interactions as a function of network distance between interacting users. Network distance is calculated in the network embedding space and then divided into quartiles, with Q1 showing the 25% of closest interactions and Q4 showing the 25% of the most distant interactions. Emotions (a) anger, (b) disgust and (c) toxicity increase with distance between users in the network embedding space, while (d) joy and (e) fear decrease with distance, as does (f) reply length. The boxes span the first to third quartiles, with whiskers extending 1.5 times the interquartile range. The horizontal line inside the box represents the median, and diamond symbol marks the mean. Differences in means were tested for statistical significance using a two-sided Mann-Whitney U Test with the Bonferroni correction: * indicates significance at p$<0.05$, ** - p$<0.01$, *** - p$<0.001$, **** - p$<0.0001$ and, ns - not-significant.