This study develops an algorithmic framework for reconstructing hidden audience votes, interpreting non-performance bias, and optimizing hybrid evaluation rules in competitive scoring systems. A multi-source vote reconstruction model integrates an industry popularity index, temporal traffic evolution, performance response, and elimination inheritance to estimate latent vote dynamics under incomplete observations. The reconstructed results reproduce historical eliminations with a consistency index of 90.66%, indicating reliable recovery of hidden preference signals. A Wilcoxon signed-rank testing framework based on a fan influence index is used to compare percentage-based and rank-based aggregation, showing that percentage aggregation can amplify extreme popularity and weaken professional judgment. An interpretable XGBoost-SHAP model quantifies external feature effects and identifies age as the dominant non-performance factor, with hometown and industry contributing weaker auxiliary influence. A time-varying convex combination model further combines Z-score normalization and sigmoid-controlled dynamic weights to shift smoothly from early-stage popularity protection to late-stage professional screening. The resulting framework improves fairness, preserves audience engagement, and provides an interpretable computational solution for evaluation mechanism optimization.
