Counterfactual Public Opinion Projection for Intervention Decision-Making

Abstract

Effective public opinion intervention requires forecasts of how the same event would evolve under alternative actions. Existing methods mainly capture observational associations and inadequately address treatment selection bias. The paper proposes a counterfactual public opinion projection framework for intervention decision making. Each intervention is represented by its timing, type, content and intensity. Causal triggering structures are inferred from propagation event sequences. Counterfactual trajectories of attention, sentiment and polarization are generated from adversarially balanced representations. Candidate interventions are then compared through repeated rollouts. Semisynthetic experiments yielded an overall F1 comparable to a direct structure learning baseline and a recall of 0.698 for causal edges linking semantically dissimilar content. Representation diagnostics showed that the gradient reversal layer reduced treatment information leakage from 0.163 to 0.140. Bootstrap analyses of real data further revealed intervention specific model effects on attention, sentiment and polarization.