SCOPE-Dyn: Latent Opinion Dynamics under Biased Sparse Observations

Abstract

Posts, votes, hyperlinks, and interaction records on online platforms reflect both users’ opinion expressions and their interaction relationships, providing an important basis for estimating users’ opinion states from platform data and analyzing their subsequent evolution. However, platform records are often incomplete and uneven: users leave visible expressions only at certain time points, so these records cannot directly form continuous and complete opinion states; moreover, active users and users with stronger opinions are often more likely to be observed, causing observed samples to systematically overrepresent such users and thereby affecting model judgments about latent opinion states. To address the problems caused by observation sparsity and selective visibility, this paper proposes SCOPE-Dyn, a latent opinion dynamics and offline graph-action scoring framework for biased and sparse social observations. The framework distinguishes latent opinions, observed expressions, and observation masks, estimates observation propensities, learns self-retention, neighbor assimilation, and opinion repulsion mechanisms, and assigns offline scores to candidate edge addition or deletion actions through forward rollouts of latent states. Controlled synthetic experiments validate model-based graph-action scoring, showing about 22% intervention-gain improvement over the strongest static baseline. Real-world experiments evaluate leakage-free one-step observable forecasting, providing predictive evidence rather than verified intervention effects on real platforms.