“What Else Are They Talking About?”: A Large-Scale Longitudinal Analysis of Misinformation Super-Spreader Communities on Facebook
•Download as PPTX, PDF•
0 likes•137 views
Paper by Daniel Angus, Stephen Harrington, Axel Bruns, Phoebe Matich, Nadia Jude, Edward Hurcombe, and Ashwin Nagappa, presented at the ICA 2024 conference, Gold Coast, 22 June 2024.
Report
Share
More Related Content
Similar to “What Else Are They Talking About?”: A Large-Scale Longitudinal Analysis of Misinformation Super-Spreader Communities on Facebook
Similar to “What Else Are They Talking About?”: A Large-Scale Longitudinal Analysis of Misinformation Super-Spreader Communities on Facebook (20)
More from Axel Bruns
More from Axel Bruns (20)
Recently uploaded
Recently uploaded (20)
“What Else Are They Talking About?”: A Large-Scale Longitudinal Analysis of Misinformation Super-Spreader Communities on Facebook
- 1. CRICOS No.00213J “What Else Are They Talking About?”: A Large-Scale Longitudinal Analysis of Misinformation Super-Spreader Communities on Facebook Prof Daniel Angus1, A/Prof Stephen Harrington1, Prof Axel Bruns1, Phoebe Matich1, Nadia Jude1, Dr Edward Hurcombe2, Dr Ashwin Nagappa1 Queensland University of Technology (QUT) Royal Melbourne Institute of Technology (RMIT)
- 3. CRICOS No.00213J Evaluating the Challenge of ‘Fake News’ and Other Malinformation • ARC Discovery Project (2020 - 2024) • Prof. Axel Bruns, Prof. Daniel Angus, A/Prof. Stephen Harrington, Dr Edward Hurcombe, Ms Jane Tan, Prof. Scott Wright (Bournemouth), Prof. Jennifer Stromer-Galley (Syracuse), Prof. Karin Wahl-Jørgensen (Cardiff) • This project conducts a systematic, large-scale, mixed-methods analysis of empirical evidence on the dissemination of, engagement with, and impact of ‘fake news’ and other malinformation in public debate, in Australia and beyond.
- 6. CRICOS No.00213J Finding Relevant Data (FakeNIX) • Iteratively updated masterlist of domains listed in existing studies of ‘fake news’ • 2,314 domains to date (from Shao et al., 2016; Starbird et al., 2017; Allcott et al., 2018; Grinberg et al., 2019; Guess et al., 2018; 2019; etc.) • Data from CrowdTangle: any Facebook posts from public pages / groups / verified profiles that contained links to any of these domains • 1 Jan. 2016 to 31 Mar. 2021: 42.6million posts from 918,760 pages/groups Limitations: • ‘Fake news’ domain lists largely US- / Anglocentric • Crowdtangle’s coverage is not complete
- 7. CRICOS No.00213J US progressives US conservatives France / Germany Italy Brazil India alternative health conspiracies UK alternative finance Nodes: public pages, groups, verified profiles / domains in posts Size: weighted in-degree Colour: weighted in-degree FakeNIX domain posts, 1 Jan. 2016 to 31 Mar. 2021 Angus, D., Bruns, A., Hurcombe, E., & Harrington, S. (2021). ‘Fake news’ on Facebook: a large-scale longitudinal study of problematic link-sharing practices from 2016 to 2020. In Selected Papers in Internet Research 2021: Research from the Annual Conference of the Association of Internet Researchers AoIR - Association of Internet Researchers. https://doi.org/10.5210/spir.v2021i0.12089
- 8. CRICOS No.00213J Thematic Mapping of Pages/Groups • Too many pages/groups in our collection to capture all content (918,760 in total) …instead… • What characterises the top 500 pages, and 500 groups in our collection? • Dimensions of interest: • the subscriber count; • count of distinct FakeNIX domains it has shared at least one link to; • number of links to any FakeNIX domain; and • total engagement from users towards this content.
- 9. CRICOS No.00213J Rank Product • Technique from genetics (Breitling et. al., 2004), useful in combining multiple quantities into a single rank. • Uses a simple geometric mean of the ranks of each quantity. • Rank product = (subscriber rank * domains rank * links rank * engagement rank) ^ 1/4 Breitling, R., Armengaud, P., Amtmann, A., & Herzyk, P. (2004). Rank products: a simple, yet powerful, new method to detect differentially regulated genes in replicated microarray experiments. FEBS letters, 573(1-3), 83– 92. https://doi.org/10.1016/j.febslet.2004.07.055
- 10. CRICOS No.00213J Analysing the ‘Prominent 1,000’ … Actually 954 • Collect all posts from these pages/groups over a six-year period: 2016 – 2021* • Text from posts (not in images) combined into a single text field • Text collated into yearly quarters per page/group: 2016Q1_1, 2016Q2_1, … , 2021Q4_954 • Latent Dirichlet Allocation (n = 40) to generate topic model for this entire corpus • Keyword analysis (top terms, bigrams, trigrams) • Qualitative exploration of emergent themes • Thematic prominence over time • Dynamic Networks of thematic similarity between pages/groups *481 pages, 473 groups = 954 total, ~70million posts
- 11. CRICOS No.00213J What (else) do they say?
- 12. CRICOS No.00213J Overall Topics • Politics • World News • Religion • Health & Wellness • Law & Order • Popular culture and entertainment • Supernatural and conspiratorial • And more…
- 13. CRICOS No.00213J Entertainment Law & Order Types of Links • Disparate sharing of link types (fake news links and non-fake news links) by different pages/groups. • Early evidence suggests this may relate to topical focus of the pages.
- 14. CRICOS No.00213J
- 15. CRICOS No.00213J Pop-culture (tv gossip, spoilers) Sports (NFL) Real-estate UK Royals Gossip Scam alerts Weather and emergency alerts
- 16. CRICOS No.00213J What else do they share?
- 17. CRICOS No.00213J Similarity Map • Combined similarity between spaces: • Non-FakeNIX domain sharing practices • Facebook on-sharing practices • YouTube video sharing practices Using practice mapping method (Bruns et al., forthcoming) • Major clusters: • Size = number of non-FakeNIX posts • Red = Trump / MAGA etc. • Blue = Berniecrats et al. • (plus other languages, other countries, evangelists, conspiracy theorists, …) Network filtered for edges with sum of cosine similarities ≥ 0.5
- 18. CRICOS No.00213J Fox News and Breitbart dominate Far more mainstream media sharing Some mainstream media sharing Biased sources, but less disinformation
- 20. CRICOS No.00213J Observations • Dominance of US political topics: • Mostly a result of US focus in FakeNIX masterlist • Lack of similar lists of problematic sources (partly because such judgments are difficult) • Divergent patterns between MAGA and Berniecrats • Berniecrats: mostly mainstream news sharing, with occasional problematic content • Trump / MAGA: mostly problematic content with occasional mainstream news (occasional content could be hate-sharing in both cases)
- 21. CRICOS No.00213J Further Outlook • Computational challenges of large datasets • Longitudinal work is still ongoing, and difficult to present in conventional paper/article formats • Large-scale topic modelling is computationally intensive and dominated by major topics (‘Trump’) • Facebook data are ill-suited to conventional network mapping practice mapping approach helps • Future repeatability: • Can we still do this with the Meta Content Library? • ‘Virtual digital enclave’ very difficult to use with standard analytics tools (Tableau, Gephi, …) • Can we use the Facebook URL Shares Dataset to extend to non-public sharing on Facebook? • Difficult with limited demographic and temporal resolution of sharing and engagement data • Can we keep the masterlist of problematic domains fresh? • Some new data sources, still very US-centric – e.g. Iffy Index of Unreliable Sources
- 22. CRICOS No.00213J This research is supported by the ARC Discovery project Evaluating the Challenge of 'Fake News' and Other Malinformation and the ARC Laureate Fellowship project Dynamics of Partisanship and Polarisation in Online Public Debate. Facebook data are provided courtesy of CrowdTangle. Acknowledgments
Editor's Notes
- Colour + size: weighted indegree