| 本文已被：浏览 2031次 下载 5465次 |
码上扫一扫！ |
|
|
| 基于概率模型检查的树模型公平性验证方法 |
|
王艳1,2, 侯哲3, 黄滟鸿1,2, 史建琦1,2, 张格林1,2
|
|
1.华东师范大学 软件工程学院, 上海 200062;2.国家可信嵌入式软件工程技术研究中心(华东师范大学), 上海 200062;3.School of Information and Communication Technology, Griffith University, Brisbane 4111, Australia
|
|
| 摘要: |
| 如今,越来越多的社会决策借助机器学习模型给出,包括法律决策、财政决策等等.对于这些决策,算法的公平性是极为重要的.事实上,在这些环境中引入机器学习的目的之一,就是为了规避或减少人类在决策过程中存在的偏见.然而,数据集常常包含敏感特征,或可能存在历史性偏差,会使得机器学习算法产生带有偏见的模型.由于特征选择对基于树的模型具有重要性,它们容易受到敏感属性的影响.提出一种基于概率模型检查的方法,以形式化验证决策树和树集成模型的公平性.将公平性问题转换为概率验证问题,为算法模型构建PCSP#模型,并使用PAT模型检查工具求解,以不同定义的公平性度量衡量模型公平性.基于该方法开发了FairVerify工具,并在多个基于不同数据集和复合敏感属性的分类器上验证了不同的公平性度量,展现了较好的性能.与现有的基于分布的验证器相比,该方法具有更高的可扩展性和鲁棒性. |
| 关键词: 公平性验证 决策树集成模型 概率模型检查 可信机器学习 |
| DOI：10.13328/j.cnki.jos.006584 |
| 分类号:TP311 |
| 基金项目:国家重点研发计划(2019YFB2102602) |
|
| Fairness Verification Method of Tree-based Model Based on Probabilistic Model Checking |
|
WANG Yan1,2, HOU Zhe3, HUANG Yan-Hong1,2, SHI Jian-Qi1,2, ZHANG Ge-Lin1,2
|
|
1.Software Engineering Institute, East China Normal University, Shanghai 200062, China;2.National Trusted Embedded Software Engineering Technology Research Center (East China Normal University), Shanghai 200062, China;3.School of Information and Communication Technology, Griffith University, Brisbane 4111, Australia
|
| Abstract: |
| More and more social decisions are made using machine learning models, including legal decisions, financial decisions, and so on. For these decisions, the fairness of algorithms is very important. In fact, one of the goals of introducing machine learning into these environments is to avoid or reduce human bias in decision-making. However, datasets often contain sensitive attributes that can cause machine learning algorithms to generate biased models. Since the importance of feature selection for tree-based models, they are susceptible to sensitive attributes. This study proposes a probabilistic model checking solution to formally verify fairness metrics of the decision tree and tree ensemble model for underlying data distribution and given compound sensitive attributes. The fairness problem is transformed into the probabilistic verification problem and different fairness metrics are measured. The tool called FairVerify is developed based on the proposed approach and it is validated on multiple classifiers based on different datasets and compound sensitive attributes, showing sound performance. Compared with the existing distribution-based verifiers, the method has higher scalability and robustness. |
| Key words: fairness verification decision tree ensemble model probabilistic model checking trustworthy machine learning |
