Under the Radar: Analyzing Recent Twitter Information Operations to Improve Detection and Removal of Malicious Actors, Part 2

Abstract

This report builds upon the work done in part one (DOI:10.2139/ssrn.4389821) of this series by examining the network structure of three information operations (IOs) that were removed from Twitter in 2021. The analysis that follows uses social network analysis (SNA) to explore the structure, key network statistics, and measures of centrality for network graphs created from Twitter mentions. Five data sets feature in this analysis, three IO networks and two control networks. Data for the three IO networks came from Twitter’s Transparency Center and contained tweets from a Russian, Chinese, and Iranian IO, respectively. In addition, two COVID-19 tweet data sets from Kaggle served as the controls. This project seeks to determine if it is possible to make cross-network comparisons that could enhance the early detection of IOs on social media platforms like Twitter. The analysis found that while each network was structurally unique, the key SNA statistics failed statistical significance testing when checking for differences between the IO group and control group. This may be the result of a small network sample size composed of a mere five networks (n=5). However, this study also found that measures of centrality had statistically significant differences between the IO group and the control group. This suggests that measures of centrality, particularly eigenvector centrality and Pagerank, could be useful metrics for differentiating IOs from legitimate Twitter conversations.

Keywords: information operations, disinformation, twitter, social media, social network analysis