Temporal signatures of coordination: detecting and anticipating influence operations on social media

Online influence operations (IOs) pose a critical threat to the integrity of public discourse on social media, yet their early detection and accurate prediction of user engagement remain open challenges. This work presents a comprehensive computational framework for understanding influence operations, addressing two interconnected questions: How can IO drivers be distinguished from legitimate users? And can future interactions between IO drivers and organic users be predicted? We analyze temporal interaction networks from five real-world influence campaigns attributed to state-sponsored actors from China, Cuba, Iran, Russia, and the United Arab Emirates. To motivate the importance of detecting IOs and predicting engagement with their content, we first provide empirical evidence that, in four of the five campaigns we study, organic user adoption of IO-pushed content (e.g., hashtags and URLs) is statistically associated with prior exposure to IO actors’ activity. Building on this evidence, we introduce two complementary modules: (i) IOBoost, a lightweight supervised classifier that leverages temporal, activity-based, and network motif features to identify IO driver accounts, achieving an average F1-score of 97.5%, demonstrating strong cross-campaign generalization and early detection capabilities for identifying IO drivers; and (ii) IOLink, a graph neural network model that leverages user interaction networks to predict future interactions between IO drivers and organic users, achieving an average ROC-AUC score of 86%. Our results reveal that coordinated influence campaigns leave distinctive behavioral traces in user adoption patterns and temporal interaction dynamics, enabling accurate and generalizable detection while remaining computationally efficient for real-world deployment. To ensure full reproducibility, we make the source code publicly available: https://github.com/SimoneMungari/IO-detecting-and-anticipating .