Algorithms and polarisation

For the culmination of my MSc in Politics & Communication, I conducted some original empirical research into the effects of attention-optimising algorithms — like Facebook’s News Feed, TikTok’s ‘For You’ or YouTube’s ‘Up Next’ recommendations — on political discourse.

Big data analysis

In the study, I analysed 10 years of Twitter data from a panel of over 1,200 politically engaged users from the UK, totalling almost 23 million tweets. I used supervised machine learning to classify which tweets were uncivil, and took Twitter’s introduction of its timeline algorithm as a discontinuity, as well as creating a retweet typology to observe algorithmic effects.

Multilevel regression analysis

Using statistical techniques like multilevel logistic regression I was able to show that the level of incivility on the platform has been increasing since it went algorithmic, estimating an increase of 42%. These findings represent a significant contribution to the understanding of social media and their fraught consequences for our politics.

Identifying user political leanings

Beyond a better understanding of social platforms and the effect they have on the media environment that the Labour Party must operate within, my dataset and methods also offer possible benefits for targeting voters online. To control for partisanship, I created an algorithm to predict a user’s partisan leaning based on the accounts they follow. As the below figures show, tested against the party in a user’s bio, it was successful. This could be used to identify persuadable voters on Twitter or elsewhere, as the technique can theoretically be transferred to other platforms.

Polarflation graph

Incivility over time among UK politically engaged Twitter users, with general trendline (left) and separate trendlines for before and after the introduction of the timeline algorithm (right).

 

Prof Robin Mansell

This is an outstanding dissertation which easily reaches High Distinction, is publishable, and exceeds expectations for the amount of work normally expected at the Master's level.

Professor of New Media and the Internet / London School of Economics

 
Social media partisanship scale
Social media partisanship and incivility graph

Distributions of predicted left–right ideology score for users based on followed political and media accounts, validated against the political party mentioned in their Twitter bio (left), and these ideology scores against user incivility, showing that more partisan users are more uncivil, except for when users follow many Labour politicians (right).

Distributions of predicted left–right ideology score for users based on followed political and media accounts, validated against the political party mentioned in their Twitter bio (top), and these ideology scores against user incivility, showing that more partisan users are more uncivil, except for when users follow many Labour politicians (bottom).