Through the looking glass: evaluating post hoc explanations using transparent models

IF 3.4 Q2 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE
Mythreyi Velmurugan, Chun Ouyang, Renuka Sindhgatta, Catarina Moreira
{"title":"Through the looking glass: evaluating post hoc explanations using transparent models","authors":"Mythreyi Velmurugan, Chun Ouyang, Renuka Sindhgatta, Catarina Moreira","doi":"10.1007/s41060-023-00445-1","DOIUrl":null,"url":null,"abstract":"Abstract Modern machine learning methods allow for complex and in-depth analytics, but the predictive models generated by these methods are often highly complex and lack transparency. Explainable Artificial Intelligence (XAI) methods are used to improve the interpretability of these complex “black box” models, thereby increasing transparency and enabling informed decision-making. However, the inherent fitness of these explainable methods, particularly the faithfulness of explanations to the decision-making processes of the model, can be hard to evaluate. In this work, we examine and evaluate the explanations provided by four XAI methods, using fully transparent “glass box” models trained on tabular data. Our results suggest that the fidelity of explanations is determined by the types of variables used, as well as the linearity of the relationship between variables and model prediction. We find that each XAI method evaluated has its own strengths and weaknesses, determined by the assumptions inherent in the explanation mechanism. Thus, though such methods are model-agnostic, we find significant differences in explanation quality across different technical setups. Given the numerous factors that determine the quality of explanations, including the specific explanation-generation procedures implemented by XAI methods, we suggest that model-agnostic XAI methods may still require expert guidance for implementation.","PeriodicalId":45667,"journal":{"name":"International Journal of Data Science and Analytics","volume":null,"pages":null},"PeriodicalIF":3.4000,"publicationDate":"2023-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"International Journal of Data Science and Analytics","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1007/s41060-023-00445-1","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
引用次数: 0

Abstract

Abstract Modern machine learning methods allow for complex and in-depth analytics, but the predictive models generated by these methods are often highly complex and lack transparency. Explainable Artificial Intelligence (XAI) methods are used to improve the interpretability of these complex “black box” models, thereby increasing transparency and enabling informed decision-making. However, the inherent fitness of these explainable methods, particularly the faithfulness of explanations to the decision-making processes of the model, can be hard to evaluate. In this work, we examine and evaluate the explanations provided by four XAI methods, using fully transparent “glass box” models trained on tabular data. Our results suggest that the fidelity of explanations is determined by the types of variables used, as well as the linearity of the relationship between variables and model prediction. We find that each XAI method evaluated has its own strengths and weaknesses, determined by the assumptions inherent in the explanation mechanism. Thus, though such methods are model-agnostic, we find significant differences in explanation quality across different technical setups. Given the numerous factors that determine the quality of explanations, including the specific explanation-generation procedures implemented by XAI methods, we suggest that model-agnostic XAI methods may still require expert guidance for implementation.
透过镜子:使用透明模型评估事后解释
现代机器学习方法允许进行复杂和深入的分析,但这些方法生成的预测模型通常非常复杂且缺乏透明度。可解释的人工智能(XAI)方法用于提高这些复杂的“黑箱”模型的可解释性,从而提高透明度并实现明智的决策。然而,这些可解释方法的固有适应性,特别是对模型决策过程的解释的忠实性,可能很难评估。在这项工作中,我们使用基于表格数据训练的完全透明的“玻璃盒”模型,检查和评估了四种XAI方法提供的解释。我们的研究结果表明,解释的保真度取决于所使用的变量类型,以及变量与模型预测之间的线性关系。我们发现每个评估的XAI方法都有自己的优点和缺点,这是由解释机制中固有的假设决定的。因此,尽管这些方法是模型不可知的,但我们发现不同技术设置的解释质量存在显着差异。考虑到决定解释质量的众多因素,包括由XAI方法实现的特定解释生成过程,我们建议与模型无关的XAI方法可能仍然需要专家指导才能实现。
本文章由计算机程序翻译,如有差异,请以英文原文为准。
求助全文
约1分钟内获得全文 求助全文
来源期刊
CiteScore
6.40
自引率
8.30%
发文量
72
期刊介绍: Data Science has been established as an important emergent scientific field and paradigm driving research evolution in such disciplines as statistics, computing science and intelligence science, and practical transformation in such domains as science, engineering, the public sector, business, social sci­ence, and lifestyle. The field encompasses the larger ar­eas of artificial intelligence, data analytics, machine learning, pattern recognition, natural language understanding, and big data manipulation. It also tackles related new sci­entific chal­lenges, ranging from data capture, creation, storage, retrieval, sharing, analysis, optimization, and vis­ualization, to integrative analysis across heterogeneous and interdependent complex resources for better decision-making, collaboration, and, ultimately, value creation.The International Journal of Data Science and Analytics (JDSA) brings together thought leaders, researchers, industry practitioners, and potential users of data science and analytics, to develop the field, discuss new trends and opportunities, exchange ideas and practices, and promote transdisciplinary and cross-domain collaborations. The jour­nal is composed of three streams: Regular, to communicate original and reproducible theoretical and experimental findings on data science and analytics; Applications, to report the significant data science applications to real-life situations; and Trends, to report expert opinion and comprehensive surveys and reviews of relevant areas and topics in data science and analytics.Topics of relevance include all aspects of the trends, scientific foundations, techniques, and applica­tions of data science and analytics, with a primary focus on:statistical and mathematical foundations for data science and analytics;understanding and analytics of complex data, human, domain, network, organizational, social, behavior, and system characteristics, complexities and intelligences;creation and extraction, processing, representation and modelling, learning and discovery, fusion and integration, presentation and visualization of complex data, behavior, knowledge and intelligence;data analytics, pattern recognition, knowledge discovery, machine learning, deep analytics and deep learning, and intelligent processing of various data (including transaction, text, image, video, graph and network), behaviors and systems;active, real-time, personalized, actionable and automated analytics, learning, computation, optimization, presentation and recommendation; big data architecture, infrastructure, computing, matching, indexing, query processing, mapping, search, retrieval, interopera­bility, exchange, and recommendation;in-memory, distributed, parallel, scalable and high-performance computing, analytics and optimization for big data;review, surveys, trends, prospects and opportunities of data science research, innovation and applications;data science applications, intelligent devices and services in scientific, business, governmental, cultural, behavioral, social and economic, health and medical, human, natural and artificial (including online/Web, cloud, IoT, mobile and social media) domains; andethics, quality, privacy, safety and security, trust, and risk of data science and analytics
×
引用
GB/T 7714-2015
复制
MLA
复制
APA
复制
导出至
BibTeX EndNote RefMan NoteFirst NoteExpress
×
提示
您的信息不完整,为了账户安全,请先补充。
现在去补充
×
提示
您因"违规操作"
具体请查看互助需知
我知道了
×
提示
确定
请完成安全验证×
copy
已复制链接
快去分享给好友吧!
我知道了
右上角分享
点击右上角分享
0
联系我们:info@booksci.cn Book学术提供免费学术资源搜索服务,方便国内外学者检索中英文文献。致力于提供最便捷和优质的服务体验。 Copyright © 2023 布克学术 All rights reserved.
京ICP备2023020795号-1
ghs 京公网安备 11010802042870号
Book学术文献互助
Book学术文献互助群
群 号:481959085
Book学术官方微信