Network-based opinion diffusion models often assume uniform influence mechanisms in diverse subpopulations. This assumption risks embedding biases against demographic minorities whose interaction patterns diverge from those of majority groups. To rigorously assess whether minorities causally affect model accuracy, we propose a computationally grounded experimental approach. Using the CoDiNG opinion diffusion model and real-world interaction data from the NetSense study, we construct counterfactual networks by systematically removing minority nodes - identified by attributes such as ethnicity, socioeconomic status, and gender - while preserving network structure through degree-preserving rewiring. Comparing model performance F1 score on original and minority excluded networks allows us to quantify the extent to which demographic minorities influence opinion diffusion.

PDF