Abstract
Detecting coordinated inauthentic behavior (CIB) typically relies on observable network connections such as retweets or financial transactions. However, as bad actors increasingly move to encrypted platforms or minimize direct interactions, these structural links become invisible. This paper proposes a novel framework that identifies coordination through "temporal fingerprints"—shared patterns of bursty activity. By analyzing inter-event times and employing a Bursty-Corrected Stochastic Blockmodel (SBM), the researchers demonstrate that temporal signals can uncover hidden communities even when network edges are entirely absent.
Introduction
The detection of coordinated behavior—ranging from market manipulation to influence campaigns—is a persistent challenge. State-of-the-art methods are often limited by their dependence on visible network topology. In the modern digital landscape, actors can easily obfuscate these links by using multiple accounts or moving to private channels. The fundamental question addressed here is: How can a community be identified when the edges of its graph are hidden? The authors argue that while links can be hidden, the timing of actions in response to external events remains a robust and difficult-to-obfuscate signal of coordination.
Methodology: Collaborative Burstiness
The proposed framework relies on three pillars:
- Inter-Event Time (IET) Analysis: Measuring the intervals between consecutive actions for individual users to capture their unique temporal signatures.
- Burstiness Similarity: Using the Kolmogorov-Smirnov (KS) statistic to compare these signatures between users, creating a similarity network independent of direct interactions.
- Bursty-Corrected SBM: A generative model extending the Stochastic Blockmodel to simulate how external "shocks" propagate through networks, triggering synchronized activity bursts.
The performance of this approach was benchmarked against structural models (e.g., node2vec) and temporal foundation models (e.g., Chronos, GPT4TS) across financial (Ethereum), social (cross-platform), and synthetic datasets.
Key Insights and the 70% Threshold
A critical finding of this research is the relationship between network visibility and detection accuracy. Structural methods, which rely on visible links, are highly effective when the network is dense. However, as visibility decreases, their performance drops sharply.
As shown in Figure 4, structural models only outperform bursty models when more than 70% of the network connections are visible. In real-world scenarios where data is fragmented or intentionally hidden, the temporal bursty signal provides a far more reliable indicator of coordination. This "latent connectivity" is harder for adversaries to hide because delaying actions to mask a bursty signature often defeats the purpose of the coordination itself, such as missing a critical trading window.
Results and Performance
The bursty model demonstrated superior performance in multi-platform settings where no direct links exist between users. In the Ethereum dataset, it achieved significantly higher Normalized Mutual Information (NMI) scores than structural methods, effectively seeing through blockchain obfuscation.
As illustrated in Figure 3, the Bursty Model consistently outperforms high-end temporal foundation models and traditional structural embeddings in identifying coordinated groups. The generative accuracy of the Bursty-Corrected SBM further validates the "shock propagation" theory by successfully reproducing heavy-tailed degree distributions observed in empirical data.
Conclusion
This research shifts the focus of coordination detection from "who interacts with whom" to "who acts together in time." By leveraging collaborative burstiness, the framework provides a powerful tool for uncovering hidden influence and manipulation across diverse digital domains. Future work will address mixed membership in overlapping communities and the development of more complex generative models to capture the full spectrum of human behavior across platforms.
Summary updated