OPTIMIZED EVALUATION OF COMPETITION SCORING MECHANISMS BASED ON UNOBSERVED DATA INVERSION AND COUNTERFACTUAL SIMULATION

Abstract

Addressing fairness disputes arising from opaque audience voting in competitive reality shows, this study proposes a quantitative evaluation framework integrating latent variable inference, mechanism performance assessment, and feature attribution analysis. First, by combining maximum entropy optimization and Bayesian inference with expert scores and elimination threshold constraints, the study reconstructs the unobservable fan vote distribution, achieving a 99.492% prediction accuracy while quantitatively assessing the confidence level of estimation results. Subsequently, a multidimensional evaluation system incorporating Audience Rescue Rate (ARR) and Technical Fairness Index (TFI) was constructed. Counterfactual simulation experiments compared the efficacy of ranking-based scoring versus proportional scoring in handling disputed cases, revealing the proportional method's pronounced tendency to amplify emotional preferences. Finally, the study introduced a Logistic-Normal share regression model to explore the differential impacts of contestants' demographic characteristics, professional backgrounds, and partner effects on evaluators. Results indicate significant aesthetic divergence between experts and audiences in social media contexts, while the partner effect explains approximately 17.37% of performance variance. This research provides scientific theoretical foundations and computational paradigms for optimizing fairness, interactivity, and professional stability in multi-criteria decision systems.