Network analysis reveals news press landscape and asymmetric user polarization

TL;DR

This study leverages a network-science framework to analyze both structural and affective polarization in online news during the 2022 Korean presidential election using a large Naver News dataset. It uncovers two opposing media groups through reaction-based signals, reveals echo chambers in co-commenting networks, and demonstrates asymmetric affective polarization in response patterns between groups. Importantly, it shows that article-level classification can be effectively performed using only collective user response statistics, achieving a macro-F1 of 0.793 that rivals text-based approaches. The work provides a generalizable, data-driven method to quantify polarization from interaction signals and offers insights into how platform dynamics shape political discourse.

Abstract

Unlike traditional media, online news platforms allow users to consume content that suits their tastes and to facilitate interactions with other people. However, as more personalized consumption of information and interaction with like-minded users increase, ideological bias can inadvertently increase and contribute to the formation of echo chambers, reinforcing the polarization of opinions. Although the structural characteristics of polarization among different ideological groups in online spaces have been extensively studied, research into how these groups emotionally interact with each other has not been as thoroughly explored. From this perspective, we investigate both structural and affective polarization between news media user groups on Naver News, South Korea's largest online news portal, during the period of 2022 Korean presidential election. By utilizing the dataset comprising 333,014 articles and over 36 million user comments, we uncover two distinct groups of users characterized by opposing political leanings and reveal significant bias and polarization among them. Additionally, we reveal the existence of echo chambers within co-commenting networks and investigate the asymmetric affective interaction patterns between the two polarized groups. Classification task of news media articles based on the distinct comment response patterns support the notion that different political groups may employ distinct communication strategies. Our approach based on network analysis on large-scale comment dataset offers novel insights into characteristics of user polarization in the online news platforms and the nuanced interaction nature between user groups.

Figures (5)

• Figure 1: Characterization of political leaning of news presses based on comment-response relation between user groups. (a) The distribution of sympathy and antipathy response per comment. (b) Sympathy ratio distribution over 50 news presses. (c) Correlation matrix based on individuals' sympathy ratio distribution. This result demonstrates that correlation between news media based on sympathy ratio uncovers two conflicting media groups with opposing ideologies.
• Figure 2: (a) Distribution of individual leanings $x$ among users. (b) Average comment activity $\left < a(x) \right >$ across different political leanings $x$. It is clearly shown that both (a) the political leaning of individual and (b) the average comment activity are peaked at the two extreme ends ($x=\pm 1$).
• Figure 3: (a) Visualization of the co-commenting interaction network, comprising $N$=21,461 users, with users' individual political leanings $x$ color-coded from blue ($x=-1$) to red ($x=+1$). Node positions were determined using a force-layout algorithm, considering only the top 2% of the total links (3,925,531 out of 355,467,509) based on link weight for visualization purposes. In the visualization, only nodes are shown. (b) and (c) Joint distributions of individual political leaning $x$ among users and the average leaning of their nearest neighbors $\left < x_{nn} \right >$ for (b) the unweighted and (c) the weighted co-commenting interaction networks. (d) and (e) show the average $\left < x_{nn} \right >$ across political leanings $x$ that corresponds to (b) and (c), respectively. The weighted version of the interaction network demonstrates a clearer separation of two clusters in (c) and higher assortativity in political leaning in (e).
• Figure 4: The relationship between the individual leaning $x$ of commenters and the corresponding responses from media groups A and B. We measure the average responses in three different ways: [(a), (d)] replies, [(b), (e)] antipathies, and [(c), (f)] sympathies for two media groups: [(a), (b), (c)] Group A and [(d), (e), (f)] Group B. While the response pattern from Group A shows a linear dependence on $x$, user responses in Group B exhibit a more mixed pattern, indicating an asymmetric response pattern between the two groups.
• Figure 5: Classification task aimed at categorizing news articles into different media groups based on response statistics (number of sympathies, antipathies, and replies). (a) Illustration of the media classification process: For each news article, the top 100 comments with the most responses were collected. The numbers of sympathies, antipathies, and replies for each comment were concatenated into a 300-dimensional vector. These vectors were used to train classification models to predict the political leaning of the news media outlet associated with each article. (b) Prediction results of the classification model using a multi-layer perceptron (MLP) on test set articles. The model was trained to output 1 for Media Group A and 0 for Media Group B. (c) Performance comparison across classification models using the macro F1-score. The first six models represent classification based on user response patterns, while the last four models are based on textual features (news headlines and full text) for comparison. The MLP achieved the highest F1-score (0.793), followed by the Random Forest model (0.790). These results highlight that distinct response patterns on comments from collective users provide meaningful signals for distinguishing media groups with different ideologies, outperforming models that rely on news headlines.