Inverse Preference Inference via Approximate Bayesian Computation and Counterfactual Simulation for Rule Evaluation in Ranking Systems

Abstract

This paper addresses the issue of unobservable audience voting data in real-world competition scenarios by proposing a unified modeling framework based on inverse inference. Treating judges' scores and elimination outcomes as observable constraints, the study constructs a Bayesian inverse problem model incorporating ranking rules by introducing a social attention prior. An approximate Bayesian computation method is employed to achieve stable estimation of the underlying voting distribution. Building upon this foundation, the paper further proposes statistical metrics to characterize result anomalies and validates the model's reliability under sparse information conditions using multi-season data. Centered on the inference results, the study establishes a counterfactual simulation and fairness evaluation system, systematically comparing the structural differences in fairness and engagement across various scoring mechanisms. Furthermore, the paper designs an automated competition mechanism requiring no subjective intervention. By employing algorithmic immunity strategies, it significantly enhances fairness while maintaining commercial activity. This method operates independently of domain-specific assumptions, making it applicable to diverse complex systems involving implicit preferences and ranking decisions. It provides a universally applicable, data-driven approach for rule design and mechanism optimization.