On Explainability in AI-Solutions A Cross-Domain Survey Simon D Duque Anton10000000340059165 Daniel

2025-05-02 0 0 362.86KB 13 页 10玖币

侵权投诉

On Explainability in AI-Solutions:

A Cross-Domain Survey?

Simon D Duque Anton1[0000−0003−4005−9165], Daniel

Schneider2[0000−0002−1736−2417], and Hans D Schotten2[0000−0001−5005−3635]

1comlet Verteilte Systeme GmbH, 66482 Zweibruecken, Germany

simon.duque-anton@comlet.de

2DFKI, 67663 Kaiserslautern, Germany

{Daniel, Hans Dieter}.{Schneider, Schotten}@dfki.de

Abstract. Artiﬁcial Intelligence (AI) increasingly shows its potential

to outperform predicate logic algorithms and human control alike. In

automatically deriving a system model, AI algorithms learn relations in

data that are not detectable for humans. This great strength, however,

also makes use of AI methods dubious. The more complex a model,

the more diﬃcult it is for a human to understand the reasoning for

the decisions. As currently, fully automated AI algorithms are sparse,

every algorithm has to provide a reasoning for human operators. For

data engineers, metrics such as accuracy and sensitivity are suﬃcient.

However, if models are interacting with non-experts, explanations have

to be understandable.

This work provides an extensive survey of literature on this topic, which,

to a large part, consists of other surveys. The ﬁndings are mapped to ways

of explaining decisions and reasons for explaining decisions. It shows that

the heterogeneity of reasons and methods of and for explainability lead

to individual explanatory frameworks.

Keywords: Artiﬁcial Intelligence ·Explainability ·Survey ·Cross-Domain.

1 Introduction

Industrial revolutions provided humanity with novel technologies that fundamen-

tally changed ﬁelds of work, mostly in manufacturing and processing industries.

Currently, the fourth industrial revolution is said to introduce ﬂexibility and

ad-hoc connectivity to once inﬂexible industrial Operational Technology (OT)

networks. This newly integrated means of connectivity in industrial networks,

?This is a pre-print of an invited paper published in the Computer Safety, Reliability,

and Security. SAFECOMP 2022 Workshops. Please cite as:

S.D. Duque Anton, D. Schneider, H.D. Schotten. On Explainability in AI-Solutions:

A Cross-Domain Survey. In: Computer Safety, Reliability, and Security. SAFE-

COMP 2022 Workshops, SAFECOMP 2022, vol 13415. Springer, September 2022,

DOI: 10.1007/978-3-031-14862-0 17

arXiv:2210.05173v1 [cs.AI] 11 Oct 2022

2 S. D. Duque Anton

across physical factory boundaries, allows for new use cases and improved ef-

ﬁciency. Apart from the connectivity aspect, the introduction of Artiﬁcial In-

telligence (AI) methods presents new paradigms. In industrial environments,

AI methods are applied to production and resource planning [27], detection of

anomalies in production processes [24,25,23], and improvement of processes [61].

Apart from industrial applications, AI methods lend themselves readily on other

domains, such as ﬁnance and banking [14], medicine [32] and elderly care [51],

autonomous driving [57,56] network management [58,39,40,22], and for control of

unmanned vehicles [55,42], just to name a few. In all of these ﬁelds, automation

and AI are intended to perform tedious and repetitive tasks to relieve work-

ers. However, this autonomous performance of tasks requires trustworthy and

understandable algorithms as an enabler. If a task is to be performed by an

algorithm, the outcome must not deviate from expectations, jitters in the input

data cannot change the outcome in an undesirable fashion. Especially regarding

AI algorithms, understanding the reasoning behind a decision is complex and

often hardly possible for human operators. This is an issue, especially regarding

regulatory standards and acceptance issues of users. Consequently, AI algorithms

need to be understandable and provide predictable outcomes in a reliable fash-

ion to further their application. This need has created the term of Explainable

Artiﬁcial Intelligence (XAI) that encompasses the need for AI methods to not

only provide sound results, but also provide the reasoning in a useful and un-

derstandable manner.

This work aims at providing an overview of requirements as well as solutions

for the explainability of AI algorithms. Works related to explainability in AI

are discussed in Section 2. Methods and techniques for explaining outcomes

of AI algorithms are introduced in Section 3. Common application scenarios

that require explainable AI methods are presented in Section 4. This work is

concluded in Section 5.

2 Related Work

This section captures an overview of related works discussing explainability of

AI methods in diﬀerent domains. A comprehensive overview is provided in Ta-

ble 1. This table lists the respective work, the domain which is discussed and

the method used to explain the decision or recommendation made by the AI

method. Reddy discusses the requirements formulated by stakeholders for ac-

ceptance of AI decisions in medical treatment and research while formulating

the point that some argue in favour of higher accuracy algorithms instead of

well-explainable ones [52]. Neugebauer et al. present a surrogate AI model, that

addresses parameter changes of the base model and consequently highlights

the relevant parameters in the decision, aiding its explainability [48]. Vilone

and Longo survey scientiﬁc research addressing XAI and categorise ﬁndings in

human-based explanations that aim at mimicking human reasoning, and objec-

tive metrics such as accuracy [64]. Caro-Mart´ınez et al. introduce a conceptual

On Explainability in AI-Solutions 3

model for e-commerce recommender systems that extends existing models with

four ontological elements: User motivation and goals, required knowledge, the

recommendation process itself, and the presentation to the user [15]. Liang et

al. present a novel online Explainable Recommendation System (ERS) that, in

contrast to commonly used oﬄine ERSs, can be updated and instantly pro-

vides explanations with recommendations [44]. Holzinger and M¨uller discuss a

mapping appraoch of explainability with causability [38]. That means creating

links between the reasoning an AI algorithm implicitely makes and the intuitive

conclusions humans draw. This is applied to the area of image-based pattern

recognition in medical treatment. Ehsan et al. present a concept for integrating

social transparency into XAI solutions [26]. They present a framework based

on expert-interviews. Angelov et al. discuss the relation of AI algorithms with

high accuracy and high explainability factors [6]. A taxonomy is provided, global

vs. local model explanation techniques for diﬀerent domains and algorithms are

surveyed and set in context with the remaining challenges. Mohseni et al. pro-

vide a survey of existing literature that clusters available methods and research

approaches respective to their design goals as well as evaluation measures [47].

Their framework is founded on the distinction between the provided categories.

Belle and Papantonis evaluate feasible methods of explainability on the use case

of a data scientist that aims to convince stakeholders [11]. Shin discusses the

relation of causability and explainability in XAI and the inﬂuence on trust and

user behaviour [59]. Singh et al. evaluate methods to explain AI conclusions

in the medical domain [60]. Barredo Arrieta et al. present an extensive survey

on literature and solutions in XAI on which they base requirements and chal-

lenges yet to conquer [9]. Ultimately, they create the concept of fair AI that

explains and accounts for decisions made. Lundberg et al. introduce a game-

theoretic model for optimal explanations in tree-based algorithms [46]. Local

explanations are combined to obtain a global explanation of the trained tree in

a human-understandable format. Ploug and Holm present the concept of con-

testable AI decision making in a clinical context [49]. Contestability means that

the decision algorithm has to provide information to the data used, any system

biases, system performance in terms of algorithmic metrics, and the decision

responsibility carried by humans or algorithms. If the decision is contested, the

algorithm has to provide insight that the alternate solution was considered as

well and has been adequately taken into account. Arya et al. introduce a col-

lection of explainability tools that are combined into a framework to provide

researchers and data scientists with the opportunity to exctract explanations

for various algorithms [8]. Linardatos et al. introduce a survey and taxonomy,

distinguishing between diﬀerent types of interpretability in AI methods before

presenting an exhaustive list of tools and methods [45]. Tjoa and Guan discuss

challenges and risks of explainability in medical AI applications [62]. They survey

existing solutions for diﬀerent algorithm types while also introducing risks and

challenges with existing solutions. Roscher et al. present an overview of methods

to conserve scientiﬁc interpretability and explainability in natural sciences [53].

Beaudoin et al. introduce a framework for explainability that can be applied in

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

OnExplainabilityinAI-Solutions:ACross-DomainSurvey?SimonDDuqueAnton1[0000000340059165],DanielSchneider2[0000000217362417],andHansDSchotten2[0000000150053635]1comletVerteilteSystemeGmbH,66482Zweibruecken,Germanysimon.duque-anton@comlet.de2DFKI,67663Kaiserslautern,GermanyfDaniel,HansDieterg.fSchneider...

展开>> 收起<<

On Explainability in AI-Solutions A Cross-Domain Survey Simon D Duque Anton10000000340059165 Daniel.pdf

共13页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

On Explainability in AI-Solutions A Cross-Domain Survey Simon D Duque Anton10000000340059165 Daniel

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: