Learning-based Priority Assignment with Preference Conflicts: An Analytical Framework and a Case Study

This paper proposes a learning-based framework for resolving preference conflicts in priority assignment by reverse-engineering task priorities from historical scheduling data. The framework integrates four modules: Conditional Generative Adversarial Network(CTGAN)-based data augmentation, Bayesian networks (BNs) for interpretable preference modeling, a novel Generative Adversarial Network(GAN)-based approach for priority learning, and constraint-relaxation optimization. This architecture systematically transforms fragmented subproblems into a integrated priority learning framework. The methodology employs differentiable proxies to extract implicit priority structures directly from sparse historical allocation patterns. umerical experiments demonstrate consistent improvements over baseline methods in scheduling accuracy. In a surgical scheduling case study, the framework attained 85.2% operating room assignment accuracy and 72.3% time slot accuracy, outperforming historical schedules by 6.17% in total reward. The results validate the framework’s effectiveness in extracting conflicting preferences and priorities from data, offering significant practical value in complex scheduling domains with resource conflicts.