A Multi-source Evaluation Fusion Model Based on MCMC Reverse Inference and DVSE Dynamic Variance Standardization

Abstract

To address the issues of inconsistency between public scores and implicit voting scales, as well as significant fluctuations in the generation of competition results, this paper investigates stable ranking methods based on the fusion of multi-source evaluation data. First, historical competition rules are transformed into computable constraints, and the distribution of undisclosed votes is estimated using a Bayesian framework and MCMC sampling. Subsequently, consistency tests and mechanism backtesting are employed to compare the stability of cumulative ranking, proportional weighting, and composite rules. The results show that the model achieves an overall consistency rate of 66.3% with historical results, with the proportion-weighted rule accounting for 73.5% of the inconsistent samples; the proposed Dynamic Variance Standardization Equilibrium (DVSE) mechanism reduces the probability of extreme controversial results by 12.4%. The study demonstrates that DVSE unifies scoring and voting scales via Z-scores, suppressing the excessive amplification of high-variance data while preserving multi-source information, thereby providing a transferable algorithmic framework for similar competition evaluation and comprehensive ranking tasks.