FIB | Networks 2021

Identifying Super-spreaders of Misinformation on Social Media

Matthew R. DeVerna, Rachith Aiyappa, Diogo Pacheco,
John Bryden, Filippo Menczer

July 7th, 2021

Online super-spreaders: a growing concern

Fake news on Twitter during the 2016 U.S. presidential election (2019) Grinberg et al., Science
The COVID-19 Infodemic: Twitter versus Facebook (2021) Yang et al., Big Data & Society

Can we identify super-spreaders of misinformation on social media?

Reliable over time

Platform agnostic

FIB index

False
Information
Broadcaster

What is the FIB index?

Repurpose the h-index...

... to capture users who consistently share low-credibility sources...

... that are reshared many times.

h- vs. FIB- index?

h-index = 100

100 publications with at least 100 citations

FIB-index = 100

100 tweets (containg a low-credibility source) with at least 100 retweets

General approach

Reconstruct a misinformation ecosystem

Query the Twitter Decahose with low-credibility sources (Iffy+)

All tweets contain a low-credibility source

10 months (Jan. – Oct. 2020)

Identify super-spreaders from the first two months

Analyze how super-spreaders account for future misinformation

General approach for testing the FIB index...

Build a misinformation ecosystem by...

Querying Twitter Decahose with a list of low-credibility domains...
We get a 10% sample of tweets containing any of these domains
10 months - Jan 2020 - Oct 2020

Identify super-spreaders from first two months using different metrics (including FIB)
See how these users account for FUTURE misinformation

Find users in the first two months, see how they account for future misinformation
Account for future misinformation well = reliable super-spreader measure

Selection metrics

All metrics calculated from the Jan/Feb misinformation data

Popularity (number of followers)
Influence (total retweets)
k-core (users from the inner most shell)
FIB index

What are the different metrics we used?
First, reminder that all metrics calculated from the jan/feb data
General concept = If you score higher on these metrics than another user, you are more of a super-spreader
Popularity - if you have more followers in the misinformation ecosystem, you are more of a super-spreader
Influence - if you got more retweets in the misinformation ecosystem, you are more of a super-spreader
kcore - if you are present in the center of the misinformation RETWEET NETWORK, you are more of a super-spreader
FIB-index - if you got a higher FIB-index in the misinformation ecosystem, you are more of a super-spreader

Which group removed the most misinformation retweets from future months?

Top FIBers and Influentials dominate

Remove +60% of the following eight months misinformation

FIBers dismantle most quickly
On average, FIBers remove ~2% more
(+490,000) misinformation retweets per user removed
in comparison to Influentials

Here is the dismantling analysis for all metrics
X-axis = Rank
Y-axis = proportion of remaining misinformation retweets
As we move from left-to-right we are removing more and more users, taking misinformation from the network each time.
Focus on the the worst offenders, based on each metric
We take the top 1% of FIB - 181 users - then take an equal amount from other groups
k-core is the exception - we take simply the central shell

Top FIBers and Influentials dominate the dismantling analysis
More than 60% of the future misinformation is removed
Removing users based on the FIB index removes the most misinformation most quickly
On average, FIBers remove ~2% more (+490,000) misinformation retweets in comparison to Influentials

How else do FIBers and Influentials differ?

How much of what they share is misinformation?

Gather all tweets sent by these users

Jan - Mar. 2020

Calculate the proportion of misinformation shared by user $i$ ($m_i$)

$$m_i = \frac{\text{# low-credibility sources shared}_i }{ \text{all sources shared}_i }$$

FIBers share ~7.4x more misinformation (per source) than influentials

Removing Influentials also removes much more non-misinformation

FIB-index is much more reliabile

What we find is that FIBers share ~7.4x more misinformation (per source) than influentials
Y-axis = m (the misinformation sharing rate)
Each dot represents an individual user
The boxplots represent the distributions where...
Red line = mean
White line = median
X-axis = type of user
Only Influentials = users only identified as a top super-spreader by relying retweet counts
Only FIBer = users only identified in the top 1% of FIB indices
Intersection = users identified by both
We can see there is a decent amount of users captured in both groups, which is what likely leads to the relatively similar performance.
However, comparing the other two groups, we can see a stark difference
Users only identified as Influentials, share misinformation much less often
As a result, by removing Influentials from the information ecosystem we are also doing damage by removing non-misinformation
FIB index is much more reliable
Relatively similar, however, the FIB index weeds out these potential fluke users

How consistent is the FIB index over time?

On average, 52% of future FIBers were previously identified in the Jan/Feb period

On average, 52% of future FIBers were previously identified in Jan/Feb period
We calculate the FIB indices for all following 2-month periods and select the top 1% of users in each time period
The number of users found is marked by the orange bar
And the blue line indicates how many of those users were already identified in Jan/Feb
We can sort of see this downward trend
However, keep in mind that even 8 months later 40% of the users identified in jan/feb are still found to be super-spreaders
Also, it may have something to do with the fact that many of these users were suspended by Twitter...

~52% of top FIBers (85 out of 181) have been suspended by Twitter

In summary...

Present a novel application of a classic network algorithm to
reliably identify super-spreaders of misinformation
on social media

FIB Index

Test it's efficacy on real-world Twitter data

What we learn...

Outperforms other tested baseline metrics

Users with the highest FIB index...

Removed +60% of future misinformation
On average, +1/5 sources shared by FIBers are low-credibility
On average, +50% of future FIBers were identified in previous months
~52% of identified users were subsequently removed by Twitter

Platform Agnostic

What we learn is...
Overall, this metric outperforms other tested metrics
The worst offenders according to this metric...

Removed 60% of future misinformation over the following eight months
Did so quite reliably considering that 1/5 of sources shared by FIBers are low-credibility
Also, it so consistently with over 50% of users being again identified in future months
Finally, ~52% had been suspended by Twitter (85/181 users)
In a way, this is our best measure because Twitter is telling us these people are breaking their rules
Last but not least, like to stress that this metric is platform agnostic because pretty much all platforms have some concept of resharing

Hopefully this can help to open the door to cross-platform analyses

@mdeverna2

mdeverna@iu.edu

Rachith Aiyappa

Diogo Pachecho

John Bryden

Filippo Menczer