2026, 39(03): 1-18

基于显性与隐性知识融合的贝叶斯多准则排序方法

A Bayesian Multi-Criteria Ranking Approach Based on the Integration of Explicit and Tacit Knowledge

DOI：10.3969/j.issn.1672-0334.2026.03.001

中国分类号：C934

作者：

刘炳胜 河北工业大学 经济管理学院，天津 300130；天津大学 管理与经济学部，天津 300072

王琦 天津大学 管理与经济学部，天津 300072

申映华 重庆大学 经济与工商管理学院，重庆 400044

陈媛 天津大学 管理与经济学部，天津 300072

基金项目：国家自然科学基金 (72134002， 72574161)；中央高校基本科研业务费专项资金 (2022CDJSKPT31-01， 2022CDJJJ-010)；重庆市自然科学基金 (CSTB2023NSCQ-MSX1075)

中文摘要：

多准则排序问题广泛存在于多源异质信息环境下的复杂决策任务中。在这类环境中，决策所依赖的知识分为以结构化数据为载体的显性知识和依托决策者经验的隐性知识，二者优势互补，如何有效整合成为提升排序质量的关键。然而，传统方法多采用串行处理架构，将两类知识分别建模、简单拼接，难以在统一框架下实现协同推理，致使排序结果的稳健性和可解释性受限。为此，在贝叶斯推理框架下提出一种融合显性知识与隐性知识的多准则排序方法，通过三阶段建模实现两类知识在同一概率空间内的交互验证。首先，采用去偏LASSO回归方法从高维准则空间中筛选统计显著的交互项，进而构建嵌入主效应和二阶交互效应的分段线性加性值优化模型，提取显性知识的结构特征并形成先验分布；其次，将决策者提供的成对比较信息构建为似然函数，量化隐性知识；最后，采用结合Hit-And-Run与Metropolis-Hastings的马尔科夫蒙特卡洛采样方法进行后验推断，刻画各方案在不同排名上的概率分布和不确定性。在5个公共数据集上，所提偏好学习模型在拟合精度和泛化性能上均显著优于已有代表性方法。在与传统排序方法的比较中，所提排序方法能够使优胜方案的概率更集中于其真实排名位置，并显著压缩中后位方案的排序区间；成对超越指数更趋向两极取值，使方案间的优劣判别边界更加清晰。上述结果表明，该方法在排序清晰度、稳健性和不确定性量化方面均明显优于传统方法。在理论上，该方法将显性知识识别出的准则交互效应系统纳入贝叶斯序数回归的先验建模，在同一概率空间内统一表征客观规律与主观偏好，缓解了已有研究中两类知识“二元割裂”的问题，拓展了多准则排序的建模范式。在实践中，该方法为应对多源异质信息环境下的复杂排序任务提供了具备稳健性和可解释性的决策支持工具，有效辅助决策者处理现实情境中的复杂偏好表达和排序判断。

关键词：多准则决策；准则交互；加性值函数；偏好学习；贝叶斯推理

英文摘要：

The multi-criteria ranking problem is pervasive in complex decision-making tasks within multi-source heterogeneous information environments. The decision-making knowledge in such scenarios can be categorized into explicit knowledge carried by structured data and tacit knowledge derived from decision-makers′ empirical experience. These two types are mutually complementary, and how to integrate them effectively becomes the key to improving ranking quality. However, traditional methods mostly adopt a serial processing architecture that models the two types of knowledge separately and combines them in a simple manner, making it difficult to achieve collaborative reasoning within a unified framework and thereby limiting the robustness and interpretability of ranking results. To address this limination, this study proposes a multi-criteria ranking method that integrates explicit and tacit knowledge within a Bayesian inference framework, realizing the interactive verification of the two types of knowledge in the same probability space through a three-stage modeling procedure. First, debiased LASSO regression is employed to screen statistically significant interaction terms from the high-dimensional criterion space, and a piecewise linear additive value optimization model embedding both main effects and second-order interaction effects is then constructed to extract the structural features of explicit knowledge and form the prior distribution. Second, the pairwise comparison information provided by the decision-maker is formulated as a likelihood function to quantify tacit knowledge. Finally, a Markov Chain Monte Carlo sampling method combining the Hit-And-Run and Metropolis-Hastings algorithms is adopted for posterior inference, characterizing the probability distribution and uncertainty of each alternative across different ranking positions. On five public datasets, the proposed preference learning model significantly outperforms representative existing methods in both fitting accuracy and generalization performance. In comparison with traditional ranking methods, the proposed ranking method concentrates the probability of leading alternatives more tightly at their true ranking positions and markedly compresses the ranking intervals of middle- and lower-ranked alternatives. Meanwhile, the pairwise outranking indices tend more strongly toward the two extremes, yielding clearer discrimination boundaries between alternatives. These results indicate that the proposed method is superior to traditional methods in ranking clarity, robustness, and uncertainty quantification. Theoretically, the method systematically incorporates the criterion interaction effects identified from explicit knowledge into the prior modeling of Bayesian ordinal regression, unifying the representation of objective regularities and subjective preferences within the same probability space, alleviating the dualistic separation of the two types of knowledge in existing research, and thereby extending the modeling paradigm of multi-criteria ranking. Practically, it provides a robust and interpretable decision-support tool for complex ranking tasks in multi-source heterogeneous information environments, effectively assisting decision-makers in handling complex preference articulation and ranking judgments in real-world scenarios.

Key words: multiple criteria decision making|criterion interaction|additive value function|preference learning|Bayesian inference