Originators vs. Amplifiers

Who is responsible for spreading misinformation on Twitter?

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

Originators and Amplifiers

  • Twitter data assumes all retweets are of the original poster
  • This renders amplifiers completely invisible

    • Who is actually driving the spread of misinformation?

  • Trumps retweet of this video shows us why this is a problem
  • API data from this tweet will attribute zero retweets to Trump

How hate and misinformation go viral: A case study of a Trump retweet
(2020, Jacob Gursky and Samuel Woolley)

General approach
  1. Gather a bunch of data
    • 4 months (Apr. — Jul. 2020)
    • Tweets contain a low-credibility source (filtered stream)
      • 100 most shared from: Shao et al. (2018), Grinberg et al. (2019), Pennycook and Rand (2019), Bovet and Makse (2019)
  1. Amplifiers analysis
    • Look for amplifiers within individual cascades (spoiler alert, we find them)
    • Measure their influence at the network level
      • Determine the best way to find superspreaders of misinformation
    • Explore behavior of amplifiers vs. originators
  1. Test how well we can predict who will become a superspreaders of misinformation
  1. Explore account status of superspreaders
    • active / suspended / verified

Reconstructing cascades

(Randomly sample 5k cascades from each month)

Probabilistic diffusion inference

Assumptions
Probability distributions

... utilized to infer how cascades actually unfolded

Two simple assumptions → two probability distributions

Follower-count probability

Users with more followers will be more likely to be retweeted

(probability proportional to number of followers)

Time-delay probability

Accounts more recently involved are more likely to be retweeted

(power-law probability distribution)

Toy example

  • Followers: a > c > b
  • Time delay (recency): c < b < a

Choosing parent of c

Choosing parent of d

This method is probabilistic...

... so we create 1,000 versions of each month's network

  • 1k reconstructed networks
  • 1 naive network
  • 1k reconstructed networks
  • 1 naive network
  • 1k reconstructed networks
  • 1 naive network
  • 1k reconstructed networks
  • 1 naive network

How can we find amplifiers?

Two criteria to define an amplifier of a cascade

domination

  • A user “dominates” cascade c when earning ≥ 𝜽 proportion of a cascade's retweets
  • 𝜽 with range [0,1] is the domination threshold

uncommon

  • Number of retweets earned in c ≥ mean_rts(c) + one_std(c)


  1. Must earn most of a cascade's retweets (defined by 𝜽)
  2. Num. retweets earned needs to be out of the ordinary

Look for amplifiers in every cascade...

...for all versions of a month's network...

...at various domination threshold levels.

Proportion of cascades with amplifiers

Substantial proportion of cascades contain amplifiers!

How influential are amplifiers within the misinformation network?

Dismantling analysis

Tested metrics

Naive + recosntructed

  1. Node strength
  2. FIB index

Reconstructed only

  1. Sum of subtree sizes*
  2. Weighted page rank
  3. Collective Influence*
  4. Optimal*

Collective influence

\[\begin{aligned} \mbox{CI}_l(i) & = ( k_i - 1 ) \sum_{j \in \partial B (i,l)} (k_j - 1) \\ \\ \partial B (i,l) & = \mbox{Ball}(i,l) \\ l & = \mbox{shortest path radius} \\ l & = 2 \end{aligned} \]

Sum of subtress

\[\begin{aligned} \mathbf{T}_i & = \sum_{c \in C_{i}} t_c\\ \\ t_c & = \mbox{size of subtree for cascade } c \\ C_{i} & = \mbox{set of cascades } i \mbox{ is invovled in} \end{aligned} \]

Optimal

Greedily select a random account from set of nodes where…

\[\begin{aligned} \mathbf{T}_i & = \max(\mathbf{T}) \end{aligned} \]

Dismantling results

Why does using the reconstructed network lead to worse results?

Percentage of amplifiers who are also originators

Originators are amplifying other originators!

Originators get more retweets than non-originator amplifiers average!

Originators are much more active than other users!

Can we predict who will become a superspreader?


  • Measure the overlapping set of top ranked users for all metrics compared to the optimal
  • Top 10, 50, 100, 1,000 users

Predicting superspreaders

Most metrics do a decent job!

What is the status of these accounts two years later?


  • Are they still tweeting?
  • Are they verified?
  • How many followers do they have?


Repulled account data for top 100 superspreaders in each month with V2 API (August 7 2022)

Account status of top 100 superspreaders

Most are still active and verified!

Number of followers by status

Accounts with more followers tend to still be active

What we learned...

  • Amplifiers do exist...
  • ...but they tend to be originators — who:
    • Are much more active than everyone else
    • Earn more retweets than non-originating amplifiers
  • Superspreaders are not hard to find
    • Tend to be verified
    • Many are still active
    • Those who are active have more followers than those who have been suspended

@mdeverna2

mdeverna@iu.edu

Rachith Aiyappa

Diogo Pachecho

John Bryden

Filippo Menczer