Bots and Their Influence During the Mexican Presidential Election – A Network Science Perspective

We know from the 2016 US election that the presence of social media bots can negatively affect democratic political discussion. This decline in civil society can potentially alter public opinion and endanger the integrity of the Presidential elections.

In February of 2017, we began collecting Twitter conversations about the ongoing elections and investigating possible interference by Twitter bots. We focused on the 2018 Mexican presidential elections by applying network analysis and machine learning tools to accurately and reliably uncover bots.

Bots and bot networks are powerful tools that can sway public opinion and set agendas for discourse. They are programmed to control online entities, create content, engage other users, and build social interactions. Bots strategically manipulate online discussions by spreading content which changes the public’s perception of important issues and the reputation of an entity, like a candidate, corporation, or campaign. They affect what we discuss and can even influence the outcome of any given political election.

How we identify Bots:
Botometer checks the activity of a Twitter account and estimates a score based on a collection of features extracted from account meta-data, social connectivity, and account activity. Higher scores indicate more autonomous behavior. Botometer is a joint project of the Indiana University Network Science Institute (IUNI) and the Center for Complex Networks and Systems Research (CNetS).

In the Mexican presidential election we analyzed more than 1 million Twitter accounts that followed the candidates and participated in relevant political discourse. Our data collection consisted of over 3.5M tweets since May 19th, 2018 and we have historical collection of tweet for different time windows. According to our analysis, bots were responsible for a significant portion of generated content. Of the accounts we inspected, 53% were bots.

We wanted to investigate how bots follow candidate accounts over time and how they create a  follower network. We applied a methodology introduced by New York Times investigation of bot markets. We ranked 100,000 randomly-selected candidate followers by recency and creation date and generated bot scores for each account in our sample. In general, all candidates have been followed by bot accounts since they were created. Human accounts follow candidates earlier in their career and when they announced their candidacy. Once the candidates announced their bid for presidency in 2018, they attracted more human followers (see depths in the middle charts) an exception Lopez Obrador, who have human followers in early phases but later attracts mostly bot followers.

How to Read “follower-inspection” Chart:
This analysis is reported in three panels:

  • Rank of follower and their corresponding creation dates which offer visual cues for anomalous behavior.
  • Mean bot score computed within a sliding window of 1,000 accounts shifting with 50% overlap to quantify the attractiveness of the account to humans and bots within a period of time.
  • Entropy of account creation date within a window to investigate irregularities in account creation dates of followers

Every Candidate has Bot Followers – For each candidate, we gathered data on two sets of 100,000 accounts:

  1. randomly-selected accounts and
  2. a subset of their most recent followers

We observed that each candidate has a vast number of bot followers, ranging between 45% and 70%. Note that the difference between bot-follower fractions is wider for the sample of recent followers, indicating that bot participation changes as the election draws near.

Human Accounts and Bots have Different Agendas – We investigated messages shared by large fractions of human and bot accounts, analyzing tweets shared more than 100 times. We investigated participation of human and bot accounts in disseminating each tweet. We found multiple anomalous tweets promoted by bots and shared around 1,000 times (see figure below).

We investigated tweets shared mostly by bot or mainly by human accounts. Based on our manual inspection of top 20 tweets in each category, we notice common patterns shared below. Also see example tweets from each category.

  • We found that accounts classified as humans tweeted pro-Ricardo Anaya and against Lopez Obrador
  • We found that accounts classified as bots tweeted pro-PRI
Human support to Ricardo Anaya
Bot fraction: 0.8% Bot fraction: 2.5% Bot fraction: 2.9% Bot fraction: 0% Bot fraction: 3.9%
Human activity against Lopez Obrador
Bot fraction: 7.2% Bot fraction: 6.6% Bot fraction: 1% Bot fraction: 0% Bot fraction: 7.7%
Bot activity supporting Jose Antonio Meade
Bot fraction: 44.4% Bot fraction: 42.7% Bot fraction: 42.6% Bot fraction: 42.1% Bot fraction: 39.3%
Bot activity against Lopez Obrador and Ricardo Anaya
Bot fraction: 67.5% Bot fraction: 39.7% Bot fraction: 28.7% Bot fraction: 68.8% Bot fraction: 29.2%

Historical Perspective of Bot Involvement

Bot Activity Observed in Early 2017 After the AHORA Announcement – We observed bot activity in early 2017 after the AHORA announcement, when a bot network was activated to support the leading candidate, López Obrador.

One example: John M. Ackerman created content to support the candidacy of López Obrador and raise concerns about the AHORA movement. In this conversation we estimate approximately 16% of the total accounts are bot-likely.

 

2017-02-17 / 2017-03-01
Nodes colored by their bot scores. Accounts participating to disseminate tweet mentioned above (red). Accounts shared across different political issues and presidential debates (in red).

20% of Accounts Still Active in 2018 for Various Campaign Related Discussions – Approximately 1,100 accounts participated in:

  • both presidential debates
  • the announcement of the #AHORA movement, and
  • discussions around Russian interference originated from Russia Today (@ActualidadRT).

20% of the accounts that participated in #AHORA discussions were also active conversation around Russian interference. We observed 4 tweets which were excessively shared that reject claims on Russian interference in Mexico.

 

2018-01-15 / 2018-01-26
Nodes colored by their bot scores. Accounts participating to disseminate relevant tweets (in red). Accounts shared across different political issues and presidential debates (in red).

Bot Network Sophistication

We Detected a Network of Bot Accounts – We inspected over 350,000 Twitter participants in the presidential debates and identified a group of accounts that were not following any of the presidential candidates. After mapping social ties of randomly selected 6,000 accounts among the the most active users, we looked for accounts that formed cliques.

  • Cliques: a group of accounts that each follow each other, forming a group of completely interconnected accounts (left image).
  • Bot networks: These cliques often coincide with bot networks (right image).

 

Examples of the most active solo accounts
For Lopez Obrador For Lopez Obrador For Lopez Obrador For Jose Antonio Meade Against Ricardo Anaya

3 thoughts on “Bots and Their Influence During the Mexican Presidential Election – A Network Science Perspective”

  1. Congratulations for this article!
    I am very surprised of how many bots are involved in the mexican elections. I’m actually mexican and Computer Science student. In fact, I was trying to do something like this for my dissertation , so I am very surprised of how do you developed the study. I am wondering if there is a journal publication about this article that digs deeply in this issue. Since I was working with a similar topic, I would like to know the metodology with more detail.
    Thank you in advance for your answer.

    Best regards,
    Eric.

  2. Interesting document, it gives a panorama of how foreign powers could potentially modify the course of an election, and it is a good experiment that is running right now. Unfortunately social media is changing and young people is using Twitter less and other tools like Whatsapp and Snapchat are becoming more common. FB is also opaque to scrutiny, but we can bet that it is also being clogged with bot accounts. Good job!

  3. Interesting article, it might have public concern due the politic actions arround the election. It is also critical how non-human participation is able to change perspective about social media tendency. As well, is important to mention that this analysis is based only in tweeter activities wich represent only a fraction of the mexican society.

Leave a Reply

Your email address will not be published. Required fields are marked *