Network Structure and Community Evolution Online: Behavioral and Emotional Changes in Response to COVID-19

Fan Fang, Tong Wang, Suoyi Tan, Saran Chen, Tao Zhou, Wei Zhang, Qiang Guo, Jian-Guo Liu, Petter Holme, Xin LüFrontiers in Public Health 2022 · DOI

TL;DR

This study leverages a massive Sina Weibo dataset from Wuhan during the early COVID-19 outbreak to analyze how social network structures, community dynamics, and user emotions evolved. They found that online activity and social connections surged during the pandemic's peak, with a significant shift in discussion topics towards COVID-19, and importantly, identified a "rebound effect" where negative posts could lead to emotional relief.

What Problem Does This Paper Address?

Understanding how societies react to large-scale, disruptive events like pandemics is crucial for effective public health responses and policy-making. While social media offers a rich source of real-time data on public sentiment and behavior, analyzing its complex dynamics during a crisis presents a significant challenge. Previous research often focused on specific aspects or timeframes. This paper aims to provide a comprehensive, multi-stage analysis of social network evolution, community shifts, and emotional changes on a major Chinese social media platform (Sina Weibo) during the initial COVID-19 outbreak in Wuhan, a critical period for understanding pandemic response.

Methods

The researchers employed a multi-faceted approach:

  • Network Construction: They built time-varying directed weighted social networks of 3.5 million Sina Weibo users. A link was established when one user mentioned or forwarded another's post.
  • Community Detection: The Infomap algorithm was used to identify evolving community structures within these networks across different stages of the epidemic. This algorithm is known for its effectiveness in finding community structures in various network types.
  • Text Analysis: Latent Dirichlet Allocation (LDA), a standard topic modeling technique, was applied to identify dominant themes and shifts in conversation topics over time.
  • Sentiment Analysis: A Bidirectional Long Short-Term Memory (Bi-LSTM) model was trained to quantify the emotional tone of microblogs, allowing for the analysis of emotional changes at both individual and population levels.

Datasets

The primary dataset used was a comprehensive collection of Sina Weibo posts from Wuhan, China, spanning from December 1, 2019, to March 20, 2020. This dataset comprised over 42 million microblogs from more than 3.5 million unique users. The sentiment analysis model was trained and validated on a manually labeled subset of 15,000 posts and then tested on a random sample of 40,000 posts (10,000 from each of the four stages).

Key Insights and Ideas

  • Amplified Online Activity: The pandemic dramatically increased online social engagement. This suggests that during crises, social media becomes an even more vital platform for communication and connection.
  • Topic Centralization: As the epidemic progressed, conversations on Weibo shifted from diverse cultural topics to a dominant focus on COVID-19, indicating a collective shift in public attention.
  • Community Cohesion: The epidemic fostered more cohesive online communities, with users increasingly discussing similar topics. This highlights how shared experiences can drive social aggregation online.
  • The "Rebound Effect" of Negative Expression: A statistically significant finding was the "rebound effect," where posting highly negative content could lead to a subsequent improvement in emotional tone. This suggests social media can serve as a cathartic outlet for processing negative emotions during stressful times.
  • Sustained Interest: A notable percentage of users (2.15% for community structure and 7.3% for user attention) maintained a consistent focus on COVID-19, even as the epidemic receded locally, possibly due to global awareness or empathy.

Key Results

  • Network Growth: During the severe stage of the epidemic, network indicators like the number of unique recipients, average degree, and transitivity increased by 24%, 23%, and 19% respectively, compared to the pre-outbreak period.
  • Topic Similarity: The similarity of discussed topics on Weibo increased significantly during the local peak of the epidemic.
  • Community Shift: The number of communities focused on COVID-19 increased by nearly 40% of the total communities.
  • Emotional Relief: A paired sample t-test revealed a statistically significant "rebound effect" (P = 0.003), indicating that negative microblog posts were followed by less negative sentiment in subsequent posts.
  • Sentiment Analysis Accuracy: The Bi-LSTM sentiment analysis model achieved accuracy rates above 86% when compared to human annotations across the four stages.

Limitations and Future Work

  • Data Completeness: The study acknowledges that some deleted microblogs might be missing, and the dataset did not actively detect or reduce misinformation.
  • User Identity Variation: The analysis did not deeply differentiate between various user types (e.g., professionals vs. general users), which could influence behavior patterns.
  • Future Directions: The authors suggest further research could explore how to better detect and mitigate misinformation, and how to differentiate user behavior based on identity variations. They also emphasize the potential of social media analytics for public health decision-making, psychological intervention, and improving government transparency.

Summary updated

Abstract

<b>Background:</b> The measurement and identification of changes in the social structure in response to an exceptional event like COVID-19 can facilitate a more informed public response to the pandemic and provide fundamental insights on how collective social processes respond to extreme events. <b>Objective:</b> In this study, we built a generalized framework for applying social media data to understand public behavioral and emotional changes in response to COVID-19. <b>Methods:</b> Utilizing a complete dataset of Sina Weibo posts published by users in Wuhan from December 2019 to March 2020, we constructed a time-varying social network of 3.5 million users. In combination with community detection, text analysis, and sentiment analysis, we comprehensively analyzed the evolution of the social network structure, as well as the behavioral and emotional changes across four main stages of Wuhan's experience with the epidemic. <b>Results:</b> The empirical results indicate that almost all network indicators related to the network's size and the frequency of social interactions increased during the outbreak. <i>The number of unique recipients, average degree</i>, and <i>transitivity</i> increased by 24, 23, and 19% during the severe stage than before the outbreak, respectively. Additionally, the similarity of topics discussed on Weibo increased during the local peak of the epidemic. Most people began discussing the epidemic instead of the more varied cultural topics that dominated early conversations. The number of communities focused on COVID-19 increased by nearly 40 percent of the total number of communities. Finally, we find a statistically significant "rebound effect" by exploring the emotional content of the users' posts through paired sample <i>t</i>-test (<i>P</i> = 0.003). <b>Conclusions:</b> Following the evolution of the network and community structure can explain how collective social processes changed during the pandemic. These results can provide data-driven insights into the development of public attention during extreme events.
BibTeX
@inproceedings{fang2022network,
  author = {Fan Fang and Tong Wang and Suoyi Tan and Saran Chen and Tao Zhou and Wei Zhang and Qiang Guo and Jian-Guo Liu and Petter Holme and Xin Lü},
  title = {{Network Structure and Community Evolution Online: Behavioral and Emotional Changes in Response to COVID-19}},
  booktitle = {Frontiers in Public Health},
  year = {2022},
  doi = {10.3389/fpubh.2021.813234},
}
Figure