摘要
超大的数据规模及结构复杂的深度模型在互联网数据的处理与应用方面表现出了优异的性能,但降低了人工智能(Artificial Intelligence, AI)系统的可解释性。反事实解释(Counterfactual Explanations, CE)作为可解释性领域研究中一种特殊的解释方法,受到了很多研究者的关注。反事实解释除了作为解释外,也可以被视为一种生成的数据。从应用角度出发,文中提出了一种生成具有高数据真实性反事实解释的方法,称为生成链接树(Generative Link Tree, GLT),采用分治策略与局部贪心策略,依据训练数据中出现的案例生成反事实解释。文中对反事实解释的生成方法进行了总结并选取了其中热门的数据集来验证GLT方法。此外,提出“数据真实性(Data Fidelity, DF)”的指标,用于评估反事实解释作为数据的有效性和潜在应用能力。与基线方法相比,GLT生成的反事实解释数据的真实性明显高于基线模型所生成的反事实解释。
The super large data scale and complex structure of deep models show excellent performance in processing and application of Internet data, but reduce the interpretability of AI systems.Counterfactual Explanations(CE) has received much attention from researchers as a special kind of explanation approach in the field of interpretability research.Counterfactual Explanations can be regarded as a kind of generated data in addition to being an explanation.From the viewpoint of application, this paper proposes an approach for generating counterfactual explanations with high data fidelity, called generative link tree(GLT),which uses a partitioning strategy and a local greedy strategy to construct counterfactual explanations based on the cases appearing in the training data.Moreover, it summarizes the generation methods of counterfactual explanations and select popular datasets to verify the GLT method.In addition, the metric of “Data Fidelity(DF)” is proposed to evaluate the fidelity and potential application of the counterfactual explanation as data from an experimental perspective.Compared with the baseline method, the data fidelity of the counterfactual explanation generated by the GLT method is significantly higher than that of the counterfactual explanation gene-rated by the baseline model.
作者
王明
武文芳
王大玲
冯时
张一飞
WANG Ming;WU Wen-fang;WANG Da-ling;FENG Shi;ZHANG Yi-fei(School of Computer Science and Engineering,Northeastern University,Shenyang 110169,China)
出处
《计算机科学》
CSCD
北大核心
2022年第9期33-40,共8页
Computer Science
基金
国家自然科学基金(62172086,61872074)。
关键词
可解释性
填充式
反事实解释
数据真实性
Interpretability
Filling type
Counterfactual explanations
Data fidelity
