Contrasting disinformation: A mixture of Hawkes Processes for the identification of coordinated inauthentic behaviour

E. Muratore${}^{a,b}$, R. Gallotti${}^b$ and C. Agostinelli${}^c$

${}^a$Università di Trento   ${}^b$Fondazione Bruno Kessler (Trento, IT)

Emerged as one of modern society’s most pressing challenges, disinformation campaigns are orchestrated to distort facts and sow societal discord. The manipulative power of these campaigns have been significantly enhanced by coordinated inauthentic behaviours (CIBs) - disguised online groups coordinated in content and timing to spread disinformation. This research investigates the detection of CIBs focusing on their anomalous temporal patterns.

We model social media activities via a mixture of Hawkes processes (HPs), stochastic processes that capture event timings and self-exciting interactions. Building on event sequence clustering methodologies [1], we opted for a Bayesian hierarchical approach using a Dirichlet distribution as the prior distribution for the $K$ mixture components, an exponential function controlling users’ influences $\varphi$, and exponential priors for each component’s parameters (${\mu }_k$, ${\alpha }_k$, ${\beta }_k$). We proceed inferring parameters and latent clustering via a Metropolis Hastings within Gibbs algorithm.

Addressing the challenge of limited ground-truth data, we first evaluate our model via a simulation framework leveraging HPs. Carefully selecting parameters, we generate synthetic CIBs mimicking different strategies: from spammers to sockpuppets (i.e. users intensely interacting to disguise themself). Our findings highlight the effectiveness of our model, obtaining high homogeneity and completeness. To further evaluate our model, we are currently experimenting on a labelled high-dimensional dataset of CIBs. Our preliminary results confirm the efficacy of our model in identifying real CIBs, paving the way to more accurate and adaptable strategies against disinformation.

Keywords: Disinformation, Mixture models, Hawkes processes