Meysam Alizadeh

Postdoctoral Research Associat

Detecting Coordinated Influence Operation Content on Social Media

  Tailored for general audience. Familiarity with supervised machine learning methods is helpful but not required.
  Friday, 03 January 2020
  09:00 - 09:45

Abstract

We study a platform-agnostic method of using available activity by coordinated influence operations on social media to detect and assess their ongoing activities. Our approach classifies the post-URL pair based on human-interpretable features without relying on user-level behavioral data. We test on data from all publicly available Twitter datasets of Chinese, Russian, and Venezuelan troll activity targeting the United States from late-2015 through 2019, and Reddit dataset of Russian influence effort during 2015 and 2016. Instead of following the conventional approach to train a classifier based on the entire dataset, we train classifiers on a monthly basis across each campaign to capture how changes in trolls activities impact the performance of our classifier over time. Prediction performances vary by month, country, platform, and experimental design, ranging from average F1 score of 0.75 to 0.94, and is robust to 1% false negative and false positive rates. Additional diagnostics test and policy implications and challenges will be discussed.

Keywords

Machine Learning Influence Operation Public Policy Democratic Elections Social Media

Bio

Computational social scientist Mesysam Alizadeh has been harnessing the wealth of network and human social data available through social media platforms to understand the roots and spread of extremist ideology. In recent projects during his Ph.D. at George Mason University and postdoctoral fellowship at Indiana University Bloomington, Alizadeh has explored the moral and emotional factors underlying political extremism. He is also studying how extremism spreads on social media by analyzing the information sharing behavior of political extremists on Twitter. Currently, Meysam is a postdoctoral research associate at the Empirical Studies of Conflict Project at Princeton University and is studying foreign influence efforts on democratic elections. In this project, advised by Professor Jacob Shapiro, Meysam is using publicly available verified data sets of foreign online influence operations to train classifiers that can identify suspicious activities on social media.