人工智能培训

搜索

机器学习论文:在机器学习分类器的反事实解释中保留因果约束(Preserving Causal Constraints in Counterfactual Exp

[复制链接]
bigrc 发表于 2019-12-9 13:06:57 | 显示全部楼层 |阅读模式
bigrc 2019-12-9 13:06:57 272 0 显示全部楼层
机器学习论文:在机器学习分类器的反事实解释中保留因果约束(Preserving Causal Constraints in Counterfactual Explanations for Machine  Learning Classifiers)解释复杂的机器学习(ML)模型的输出通常需要使用更简单的模型进行近似。为了构造也与原始ML模型一致的可解释的解释,提出了反事实示例-显示模型的输出如何在对输入进行较小扰动的情况下发生变化--。本文通过解决此类示例的可行性挑战,扩展了反伪平面的工作。对于在诸如医疗保健,金融等关键领域中的ML模型的解释,反事实示例仅在特征输入微扰在现实世界中可行的情况下才对最终用户有用。我们将可行性问题表述为保留输入要素之间的因果关系,并提出一种使用(部分)结构因果模型来生成可操作的反事实的方法。当可行性约束可能难以表达时,我们提出了一种替代方法,当人们与其输出进行交互并提供类似于oracle的反馈时,该方法针对可行性进行了优化。我们在贝叶斯网络和广泛使用的“成人”数据集上的实验表明,我们提出的方法可以生成满足可行性约束的反事实解释。
Explaining the output of a complex machine learning (ML) model often requiresapproximation using a simpler model.To construct interpretable explanationsthat are also consistent with the original ML model, counterfactual examples--- showing how the model's output changes with small perturbations to theinput --- have been proposed.This paper extends the work in counterfactualexplanations by addressing the challenge of feasibility of such examples.Forexplanations of ML models in critical domains such as healthcare, finance, etc,counterfactual examples are useful for an end-user only to the extent thatperturbation of feature inputs is feasible in the real world.We formulate theproblem of feasibility as preserving causal relationships among input featuresand present a method that uses (partial) structural causal models to generateactionable counterfactuals.When feasibility constraints may not be easilyexpressed, we propose an alternative method that optimizes for feasibility aspeople interact with its output and provide oracle-like feedback.Ourexperiments on a Bayesian network and the widely used "Adult" dataset show thatour proposed methods can generate counterfactual explanations that satisfyfeasibility constraints.机器学习论文:在机器学习分类器的反事实解释中保留因果约束(Preserving Causal Constraints in Counterfactual Explanations for Machine  Learning Classifiers)
URL地址:https://arxiv.org/abs/1912.03277     ----pdf下载地址:https://arxiv.org/pdf/1912.03277    ----机器学习论文:在机器学习分类器的反事实解释中保留因果约束(Preserving Causal Constraints in Counterfactual Explanations for Machine  Learning Classifiers)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

bigrc当前离线
新手上路

查看:272 | 回复:0

快速回复 返回顶部 返回列表