| 本文已被：浏览 2133次 下载 5639次 |
码上扫一扫！ |
|
|
| 基于最小不满足核的随机森林局部解释性分析 |
|
马舒岑1,2, 史建琦1,2, 黄滟鸿1,2, 秦胜潮3, 侯哲4
|
|
1.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;2.华东师范大学 软件工程学院, 上海 200062;3.深圳大学计算机与软件学院, 广东 深圳 518060;4.School of Information and Communication Technology, Griffith University, Brisbane 4111, Australia
|
|
| 摘要: |
| 随着机器学习在安全关键领域的应用愈加广泛,对于机器学习可解释性的要求也愈加提高.可解释性旨在帮助人们理解模型内部的运作原理以及决策依据,增加模型的可信度.然而,对于随机森林等机器学习模型的可解释性相关研究尚处于起步阶段.鉴于形式化方法严谨规范的特性以及近年来在机器学习领域的广泛应用,提出一种基于形式化和逻辑推理方法的机器学习可解释性方法,用于解释随机森林的预测结果.即将随机森林模型的决策过程编码为一阶逻辑公式,并以最小不满足核为核心,提供了关于特征重要性的局部解释以及反事实样本生成方法.多个公开数据集的实验结果显示,所提出的特征重要性度量方法具有较高的质量,所提出的反事实样本生成算法优于现有的先进算法;此外,从用户友好的角度出发,可根据基于反事实样本分析结果生成用户报告,在实际应用中,能够为用户改善自身情况提供建议. |
| 关键词: 机器学习可解释性 特征重要性 反事实样本 形式化方法 逻辑推理 |
| DOI：10.13328/j.cnki.jos.006586 |
| 分类号:TP311 |
| 基金项目:国家重点研发计划(2019YFB2102602) |
|
| Minimal-unsatisfiable-core-driven Local Explainability Analysis for Random Forest |
|
MA Shu-Cen1,2, SHI Jian-Qi1,2, HUANG Yan-Hong1,2, QIN Sheng-Chao3, HOU Zhe4
|
|
1.National Trusted Embedded Software Engineering Technology Research Center (East China Normal University), Shanghai 200062, China;2.Software Engineering Institute, East China Normal University, Shanghai 200062, China;3.College of Computer Science and Software Engineering, Shenzhen University, Shenzhen 518060, China;4.School of Information and Communication Technology, Griffith University, Brisbane 4111, Australia
|
| Abstract: |
| With the broader adoption of machine learning (ML) in security-critical fields, the requirements for the explainability of ML are also increasing. The explainability aims at helping people understand models’ internal working principles and decision basis, which adds their realibility. However, the research on understanding ML models, such as random forest (RF), is still in the infant stage. Considering the strict and standardized characteristics of formal methods and their wide application in the field of ML in recent years, this work leverages formal methods and logical reasoning to develop a machine learning interpretability method for explaining the prediction of RF. Specifically, the decision-making process of RF is encoded into first-order logic formula, and the proposed approach is centered around minimal unsatisfiable cores (MUC) and local interpretation of feature importance and counterfactual sample generation method are provided. Experimental results on several public datasets illustrate the high quality of the proposed feature importance measurement, and the counterfactual sample generation method outperforms the state-of-the-art method. Moreover, from the perspective of user friendliness, the user report can be generated according to the analysis results of counterfactual samples, which can provide suggestions for users to improve their own situation in real-life applications. |
| Key words: explainable machine learning feature importance counterfactual sample formal method logical reasoning |
