Interactions among Echo Chambers in Social Media

Team: Jiazan Shi, Siming Wu, Yue Yang(Max)

Sponsor: Maurizio Porfiri, Rayan Succar, Alain Boldini

Project Objective

Objective

Hypothesis

• Users tend to following like-minded users —— echo chamber
• Influential nodes have more connections to other communities

Echo chamber refers to the situation where opinions get amplified and reinforced by information spread inside a closed system because people tend to believe the information that is similar or could support their own beliefs. The effect of echo chambers could be more severe when it comes to political and other sensitive topics, such as gun control, which would cause polarization and social danger.

• Structure of echo chambers on Twitter
• Features of influential nodes

2

Literature Review: Echo Chamber

• Echo chamber is users show tendency to favor like-minded information and join groups formed around a shared narrative
• According to group polarization theory, an echo chamber can act as a mechanism to reinforce an existing opinion within a group and move the entire group toward more extreme position
• There are homophilic clusters on social media discussion, that is users are surrounded by like-minded users. Social medias like Facebook and Twitter show polarity than social medias with a feed algorithm tweakable by users
• Different type of interactions and different platform may show different results. One could assume reply and mention show cross-cutting interactions, while following show more political homophily

3

Literature Review: Stance Detection

Lexicon Based Methods

• Create an arguing lexicon that combine arguing and sentiment features
• Accuracy = 63.93%

ML Based Methods

• Apply machine learning models, SVM or Naive Bayes, for stance classification
• SVM with ngram: F1 score = 68.98

​

DL Based Methods

• Use deep learning models, like RNN or Bert
• RNN: F1 score = 67.82

Pros

Cons

• Provide fair performance

• Require large dataset
• Easy to overfitting

4

Data Source

Vaccine:

• 02/28/2021 - 03/15/2021
• Covid-19 Vaccines

• Vaccine
• Vaccination
• Vaccinate
• Pro vax
• Anti vax

Abortion:

• 06/17/2022 - 07/01/2022
• Roe v. Wade Overturn

• Abortion
• Roe wade
• Interrupt pregnancy
• Planned parenthood
• Pro choice
• Pro life

Gun:

• 05/17/2022 - 05/31/2022
• Mass Shootings

• Gun
• Firearm
• Assault weapon
• Second amendment
• NRA
• Self defence

5

How did we choose scraping period?

— Google trends

Gun

Vax

Abortion

Covid-19 Vaccines

School mass shooting

Roe v Wade overturned

6

Data Pipeline

Tweets Scraping

(Snscrape)

Users Following

(Twint)

Cleaning

&

Filtering

Subset of Tweets

Stance Detection

(ChatGPT)

Network Analysis

User Following Relationship

Active Users

Hypothesis Testing

Visualization

Stance Leaning

Sentiment

Analysis

​

Network of

Pro and Anti Users

Data Collection

Data Preprocessing

Data Analysis

7

Data Summary

8

Methodology: Concept & Definition

• Following relationship among users

• Stance Score: stance label from GPT
• Sentiment Score: Polarity score [-1, 1]
• Tweet leaning = stance*abs(sentiment)

• the average leaning of a series of tweets produced by the user

• the average leaning of its connected nodes

9

Methodology: ChatGPT

Model:

gpt-3.5-turbo

Prompt:

“Is this a pro-gun, anti-gun, neutral, or off-topic statement? Only say it is neutral when it is completely non-opinionated: …”

Response:

“This tweet is anti-gun”

ChatGPT Validation

10

Methodology: Analysis

11

Analysis: Network Structure

Abortion

Gun

Vaccine

12

Analysis: Echo Chamber

Corr = 0.478

P-value = 0.00

Abortion

Gun

Vaccine

Corr = 0.04

P-value = 0.00

Corr = 0.455

P-value = 0.00

13

Analysis: Cross-community Interaction

Corr = 0.141

P-value = 0.00

Corr = 0.217

P-value = 0.00

Corr = 0.116

P-value = 0.00

14

Analysis: Cross-stance Interaction

Corr = 0.185

P-value = 0.00

Corr = 0.083

P-value = 0.00

Pro-gun corr = 0.363

Anti-gun corr = 0.101 (limited)

Pro-abortion corr = 0.176

Anti-abortion corr = 0.185

Pro-vax corr = 0.159

Anti-vax corr = 0.116

Corr = 0.129

P-value = 0.00

15

Conclusion

• Echo chamber exist for Abortion and Vaccine topics, can not conclude that for Gun

​

​

• Influential nodes have more connections to other communities and to members of other stance

16

Findings

• cross-community Interaction Corr Gun > Abortion > Vaccine. Gun users are more willing to follow people of another stance, vaccines are the least willing.

• During the "Robb Elementary School shooting" period, pro-gun voices were much quieter, with anti-gun users outnumbering them 15:1

• Both of the abortion and vaccine topics have weaker correlation between indegree and cross-communities out-degree than correlation between indegree and cross-stance out-degree

17

Findings

• Discussion volume varies by topic. Given that guns are integral to U.S. culture, avg number of tweets per user – guns (2.39) > vaccines (2.19) > abortion (1.99)

​

• Vaccine discussions are balanced, while anti-abortion (2:1) and anti-gun(15:1) sentiments dominate respective topics.

18

Limitation

• There are likely better ways to take into account sentiment score in Stance classification than direct multiplication

​

• The minimum and maximum tweet count threshold excluded potentially interesting users

​

• We only used following relations as network edges, other candidates include reply/retweets/like

​

• Chagpt 3.5 is very cautious regarding giving a stance resulting in many neutral labels.

​

​

19

Stance detection, is there ground truth?

​

​

​

A person's a person, no matter how small. - Dr. Suess #WAAR

​

​

​

Thanks for gently spoken truth, @DanaPerino ! How DO we as a nation "protect the innocent"? #TheFive #CCOT #DNC #RNC

​

​

​

I can tell women what to do with their bodies #ThingsYouDontSayAsAPolitician

​

​

​

20

Stance detection, is there ground truth?

​

​

​

A person's a person, no matter how small. - Dr. Suess #WAAR

​

​

​

Thanks for gently spoken truth, @DanaPerino ! How DO we as a nation "protect the innocent"? #TheFive #CCOT #DNC #RNC

​

​

​

I can tell women what to do with their bodies #ThingsYouDontSayAsAPolitician

​

​

​

21

Policy Implications

Promote diverse viewpoints in algorithm: For sensitive or political information, we suggest algorithm engineers to reduce the weight for similarity and include diversity when designing algorithms.

Increase the connections between communities for sensitive communities: Twitter applies SimClusters to detect communities. Once the communities are identified, we encourage them to increase the recommendations from the opposite communities.

Allow users to personalize their feed option: A previous paper showed social media platforms that don’t have feed algorithms tweakable by users have higher possibility to have echo chamber effect than platforms that have tweakable algorithms.

Surveillance for online discussion？Some suggest that the government should increase surveillance on online discussion of sensitive or political topics, what do you think?

22

Community Notes

• A collaborative way to add helpful context to Tweets and keep people better informed.
• Users sign up to be contributors.
• Contributors write and rate notes.
• Notes that rated as helpful by people who come from different perspectives are displayed.

Currently, very few tweets get notes. Can it be scaled?

23

Advanced cluster detection techniques are in hand

• SimClusters discover communities led by influential users
• 145k communities, from hundreds to millions of users each
• Tweets belong to one or more community

24

References

• Cinelli, M., De Francisci Morales, G., Galeazzi, A., Quattrociocchi, W., & Starnini, M. (2021). The echo chamber effect on social media. Proceedings of the National Academy of Sciences, 118(9), e2023301118.
• Lewandowsky, S., Ecker, U. K., Seifert, C. M., Schwarz, N., & Cook, J. (2012). Misinformation and its correction: Continued influence and successful debiasing. Psychological science in the public interest, 13(3), 106-131.
• Barberá, P. (2020). Social media, echo chambers, and political polarization. Social media and democracy: The state of the field, prospects for reform, 34.
• Barberá, P. (2015). Tweeting from left to right: Is online political ... - sage journals.
• Cinelli, M. (2021). The Echo Chamber Effect on social media | PNAS.
• Grömping, M. (2014). “echo chambers” partisan Facebook groups during the 2014 Thai election. Scinapse.
• Williams, H. T. P., McMurray, J. R., & Kurz, T. (2015, April 11). Network analysis reveals open forums and Echo Chambers in social media discussions of climate change. Global Environmental Change.
• Somasundaran, S., & Wiebe, J. (2010). Recognizing stances in ideological on-line debates - ACL anthology.
• Mohammad, S. M., Kiritchenko, S., Sobhani, P., Zhu, X., & Cherry, C. (2016). Semeval-2016 task 6: Detecting stance in Tweets - ACL Anthology.
• Ghosh, S., Singhania, P., Singh, S., Rudra, K., & Ghosh, S. (2019). Stance detection in web and Social Media: A comparative study. Lecture Notes in Computer Science, 75–87.
• Twitter. (n.d.). Twitter’s recommendation algorithm. Twitter.
• Sobolevsky, S., Campari, R., Belyi, A., & Ratti, C. (2014, October 1). A general optimization technique for high quality community detection in complex networks. arXiv.org.

25