Individual Belief Networks from Online Comments: Construction, Validation, and Comparative Analysis

Belief networks are a useful representation of how interrelated political opinions form coherent systems in the mind. However, prior studies have largely relied on static, population-level analysis, failing to capture the dynamic, idiosyncratic structures of individual belief networks over time. Here we present a large-scale, longitudinal analysis of individual-level reconstructed belief networks, drawing from the commenting histories of 300 users across three politically divergent news platforms over multiple years. Using a state-of-the-art large language model validated against human gold standards, we extract stances on 63 political topics across a corpus of 3.15 million comments. We report three findings. First, the reconstructed belief networks are interpretable, distinguish users who take the same stances, and are specific to the user who wrote them. Most of each network’s structure is individual and does not reduce to a few shared types. Second, users of the same platform agree to different degrees: Breitbart users agree more with each other than Mother Jones users, both on their stances and on how they link topics. The Hill, a centrist platform, has no centrist network: its users split into 29 left-leaning and 71 right-leaning users, whose networks resemble those of Mother Jones and Breitbart users even though they comment on a different platform. Third, belief networks change as political events unfold: after his candidacy, Trump becomes the most connected topic of most networks, one that most right-leaning users favor and left-leaning users are against. In conclusion, by tracking individual belief networks over time, we offer a granular view of belief dynamics at scale, mapping how individual users organize their positions and converge on shared central topics within large, polarized online communities. To facilitate continued research and build upon these insights, we make the yearly belief networks of all 300 users publicly accessible, and will release the data, an interactive viewer, and the code of our methodological pipeline upon publication.