Content-based features predict social media influence operations.

Area of Science:

• Social Media Analysis
• Computational Social Science
• Machine Learning

Background:

• Coordinated influence operations on social media pose a threat to information integrity.
• Distinguishing these operations from organic activity is crucial for platform security and democratic processes.
• Existing methods often rely on platform-specific data or complex feature sets.

Purpose of the Study:

• To evaluate a platform-agnostic machine learning approach for detecting social media influence operations.
• To assess the effectiveness of human-interpretable, content-derived features in identifying coordinated campaigns.
• To determine if these methods can differentiate influence operations from organic user activity.

Main Methods:

• A machine learning model was developed using publicly available Twitter and Reddit data.
• The model focused on human-interpretable features extracted solely from user-generated content.
• Classifiers were trained and tested monthly on data from Chinese, Russian, and Venezuelan influence campaigns targeting the US.

Main Results:

• Content-based features effectively distinguished influence operations from organic social media activity.
• The approach demonstrated consistent performance across different time periods, countries, platforms, and prediction tasks.
• Industrialized content production in influence campaigns leaves a detectable and trackable signal.

Conclusions:

• A content-focused, platform-agnostic machine learning method can reliably detect and track social media influence operations.
• Human-interpretable features derived from content are sufficient for identifying coordinated inauthentic behavior.
• This approach offers a scalable solution for monitoring and mitigating online influence campaigns.