主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

中国管理科学 ›› 2026, Vol. 34 ›› Issue (9): 79-89.doi: 10.16381/j.cnki.issn1003-207x.2024.1896

• • 上一篇    下一篇

考虑情境依赖决策行为的多准则决策偏好学习方法

梁倩1, 扈衷权2, 张震3(), 王浩旻1   

  1. 1.西南财经大学管理科学与工程学院,四川 成都 611130
    2.西安电子科技大学经济与管理学院,陕西 西安 710126
    3.大连理工大学经济管理学院,辽宁 大连 116024
  • 收稿日期:2024-10-21 修回日期:2025-03-10 出版日期:2026-09-25 发布日期:2026-09-01
  • 通讯作者: 张震 E-mail:zhen.zhang@dlut.edu.cn
  • 基金资助:
    国家自然科学基金项目(72371049);国家自然科学基金项目(71971039);国家自然科学基金项目(72201215);教育部人文社会科学基金项目(23YJC630219);四川省自然科学基金项目(2023NSFSC1025)

Multi-Criteria Preference Learning Method Considering Context-Dependent Decision Behavior

Qian Liang1, Zhongquan Hu2, Zhen Zhang3(), Haomin Wang1   

  1. 1.School of Management Science and Engineering,Southwest Finance and Economics,Chengdu 611130,China
    2.School of Economics and Management,Xidian University,Xi'an 710126,China
    3.School of Economics and Management,Dalian University of Technology,Dalian 116024,China
  • Received:2024-10-21 Revised:2025-03-10 Online:2026-09-25 Published:2026-09-01
  • Contact: Zhen Zhang E-mail:zhen.zhang@dlut.edu.cn

摘要:

在多准则决策分析中,为降低决策者的认知负担,可能会要求决策者提供关于参考方案的偏好信息,以学习决策者对方案评价的效用函数,这一问题称作偏好学习问题。本文针对多准则决策问题,提出了一种考虑决策者情境依赖决策行为的偏好学习方法。利用该方法,首先通过构建优化模型检验决策者提供的偏好信息的一致性,进而判断重构决策者提供的偏好信息所需要的效用函数数量。在此基础上,本文遵循偏好学习的简约性原则,构建优化模型,识别出一组互补的效用函数来重构决策者的偏好信息,并采用分类与回归树算法识别不同准则情境下决策者选择效用函数的规则。最后,本文以科研机构评估问题为例,通过算例分析、对比分析和仿真分析,验证了所提偏好学习方法的可行性和有效性。

关键词: 多准则决策, 情境依赖行为, 效用函数, 偏好学习

Abstract:

Multi-criteria decision analysis (MCDA) often requires decision makers (DMs) to provide preference information for a subset of reference alternatives, thereby reducing their cognitive burden while capturing their underlying value functions. When a DM’s holistic preference information is inconsistent with an assumed model, either some assignments cannot be reproduced, or the DM must revise their preferences. In this study, a specific challenge is addressed: the inconsistency arising from criterion context-dependent decision behavior. As DMs gather more information about alternatives, they may adopt different evaluation strategies across various criteria contexts, leading to inconsistencies in their preference information. To tackle this issue, the DM’s preferences are modeled using a piecewise linear value function and develop a two-stage methodological framework. In the first stage, an optimization model is constructed to assess the consistency of the provided preference information and to determine the number of value functions required for an accurate reconstruction. In the second stage, leveraging the principle of parsimony, an optimization model is introduced to select a complementary set of value functions that best reconstruct the DM’s preference information. Furthermore, classification and regression tree (CART) algorithms are integrated to extract rules that correlate different criteria contexts with the corresponding value functions, thereby enhancing the interpretability and decision support of the model. The proposed approach is validated through a comprehensive case study on the evaluation of research units. Detailed numerical experiments, comparative assessments, and simulation analyses demonstrate that our method effectively captures the context-dependent decision behavior of DMs, while also improving the robustness and explanatory power of the preference model.

Key words: multiple-criteria decision-making, context-dependent decision behavior, value function, preference learning

中图分类号: