Physician ranking based on heterogeneous medical user-generated content: A Bayesian preference disaggregation model considering criteria interactions
本研究融合自然语言处理与贝叶斯偏好分解模型，从医疗平台用户评论中提取异构决策信息，考虑准则间的交互作用，实现对医生的有效排名。
This study develops a data-driven framework integrating natural language processing (NLP) and a Bayesian preference disaggregation model to rank physicians using user-generated content (UGC) on medical platforms. Previous studies failed to extract diverse decision-making criteria and corresponding values accurately and neglected criteria interactions in extracting user preferences. To address these gaps, a pre-trained Chinese-RoBERTa model is fine-tuned to accurately compute sentiment polarities from text reviews, while a qualifier corpus is developed to assess sentiment intensities. The extracted decision information includes crisp values, interval values, and probabilistic linguistic term sets. To model this heterogeneous information, a Bayesian preference disaggregation model is designed to disaggregate pairwise comparisons of alternatives, deriving value functions that capture both marginal and interaction utilities of criteria. The Metropolis-Hasting algorithm is employed to solve the model, and the inferred value functions are used to rank physicians. A case study utilizing physician reviews from haodf.com demonstrates the effectiveness of the proposed framework, with comparative analysis of two case studies against Bayesian ordinal regression and stochastic ordinal regression methods further validating its performance.