Detecting Influence Campaigns in Social Networks Using the Ising Model

We consider the problem of identifying coordinated influence campaigns conducted by automated agents or bots in a social network. We study several different Twitter datasets which contain such campaigns and find that the bots exhibit heterophily - they interact more with humans than with each other. We use this observation to develop a probability model for the network structure and bot labels based on the Ising model from statistical physics. We present a method to find the maximum likelihood assignment of bot labels by solving a minimum cut problem. Our algorithm allows for the simultaneous detection of multiple bots that are potentially engaging in a coordinated influence campaign, in contrast to other methods that identify bots one at a time. We find that our algorithm is able to more accurately find bots than existing methods when compared to a human labeled ground truth. We also look at the content posted by the bots we identify and find that they seem to have a coordinated agenda.

💡 Research Summary

The paper tackles the problem of detecting coordinated influence campaigns on social media, specifically focusing on automated accounts (bots) that operate on Twitter. The authors begin by collecting six diverse Twitter datasets covering political events and social movements in the United States, France, and Hungary, amassing between one and three million tweets and up to one million users per dataset. For each event, roughly 300 accounts were manually labeled as human, bot, or “no idea,” yielding an overall bot prevalence of about 10 %. To enlarge the ground‑truth set, they also applied an existing bot‑detection tool (Davis et al., 2016) and used a 0.5 probability threshold.

A key empirical observation is that bots display heterophily: they retweet humans far more often than they retweet other bots, whereas humans tend to retweet other humans. The authors quantify this by computing “retweet rates” (average number of retweets per target type) and performing Kolmogorov‑Smirnov tests; all p‑values are below 10⁻⁶, confirming statistically significant differences between the four possible interaction types (B→H, B→B, H→H, H→B). This heterophily is the cornerstone of their modeling approach.

The authors formulate a probabilistic graphical model that mirrors the Ising model from statistical physics. Each node i in the interaction graph G = (V,E) carries a latent binary variable Δᵢ (0 = human, 1 = bot). Node features xᵢ contribute a unary energy φ(xᵢ,Δᵢ), while each directed edge (i→j) contributes a pairwise energy ψ(zᵢⱼ,Δᵢ,Δⱼ). The pairwise energy depends on the number of retweets wᵢⱼ, the out‑degree of the source (zᵢ), and the in‑degree of the target (zⱼ). Specifically:

ψᵢⱼ = wᵢⱼ · γ ·