Detecting and analyzing bots on Finnish political twitter

Abstract

This master’s thesis develops a machine learning model for detecting Twitter bots and applying the model to assess if bots were used to influence the 2019 Finnish parliamentary election. The aim of the thesis is to contribute to the growing information systems science literature on the use of social media and information systems to influence voters as well as to increase the general awareness in Finland of the effects of bots on Twitter.

The thesis relies primarily on quantitative analysis of a dataset consisting of 550,000 unique Twitter accounts. The data was collected from Twitter during March 2019. The accounts in the dataset belong to humans and bots that were following 14 prominent Finnish politicians on Twitter. To determine which accounts are bots and to assess the feasibility of a new method for Twitter bot detection, a machine learning model that utilizes metadata-based features for classifying Twitter accounts as bots or humans is developed and tested on the dataset.

The findings of this thesis indicate that a metadata-based approach is suitable for detecting bots and that there are several large botnets in the Finnish Twittersphere. Over 30% of the 550,000 accounts are labeled as bots by the model, which implies that the prevalence of bots is much higher than previously suggested by Twitter’s official estimates. Furthermore, a majority of the accounts seem inactive and either no longer being used or dormant and waiting for activation. The purpose of most of the bot accounts is obscure, and it is not certain how many of them are following and inflating the politicians’ popularity on purpose. Although the bots clearly increase the visibility of certain politicians, the effects of the bots on Finnish political Twitter are deemed negligible.