=Paper= {{Paper |id=Vol-3017/xcbr97 |storemode=property |title=Evaluating CBR Explanation Capabilities: Survey and Next Steps |pdfUrl=https://ceur-ws.org/Vol-3017/97.pdf |volume=Vol-3017 |authors=Lawrence Gates,David Leake |dblpUrl=https://dblp.org/rec/conf/iccbr/GatesL21 }} ==Evaluating CBR Explanation Capabilities: Survey and Next Steps== https://ceur-ws.org/Vol-3017/97.pdf

Evaluating CBR Explanation Capabilities:
Survey and Next Steps

Lawrence Gates and David Leake

Luddy School of Informatics, Computing, and Engineering, Indiana University
Bloomington IN 47408, USA
gatesla@indiana.edu, leake@indiana.edu

Abstract. The promise of case-based reasoning (CBR) methods for ex-
plaining intelligent systems has prompted much interest, including inter-
est in explaining the CBR process itself. However, there is an impediment
to understanding and advancing explanations of the CBR process: There
has been little assessment of how well they perform. This paper develops
strategies for ﬁlling this gap. We ﬁrst highlight selected work on eval-
uating systems in explainable artiﬁcial intelligence, and then present a
selective survey of work on explainable CBR systems to identify those
that have been assessed with human subjects studies. From this we de-
velop a proposed set of dimensions for characterizing explanation com-
ponents of CBR systems. These dimensions provide a framework that
can be used to guide evaluation, as well as a vehicle for identifying new
research questions. We close with proposed next steps for research and
for community initiatives.

Keywords: Explainable Case-based Reasoning, Explanation, Evalua-
tion, Human Subjects, Survey

1 Introduction

Explanation of intelligent systems is a major area of AI research. Progress has
been driven both by AI research initiatives, most notably DARPA’s Explainable
Artiﬁcial Intelligence (XAI) Program [17], and by practical needs arising from
the European Union directive on the “right to explanation” for conclusions from
decision-making systems [16]. The CBR community has long seen interpretabil-
ity as a key advantage of CBR (e.g., [25, 27]). There is substantial evidence for
human use of case-based reasoning [26], and experience with CBR applications
suggests that people ﬁnd prior examples a natural vehicle for explaining de-
cisions. Numerous approaches have been proposed for using CBR to generate
explanations for black box systems (e.g., [20, 21, 36]), including “twinning” CBR
and black box systems and using cases to explain the black box system results
(see Keane and Kenny [20] for a survey). There are also strong arguments for
focusing directly on the construction of interpretable systems rather than at-
tempting to augment noninterpretable ones with explanatory capabilities [44]—
another opportunity for CBR. Workshops on explainable CBR (XCBR) [2, 6]

Copyright © 2021 for this paper by its authors. Use permitted under
Creative Commons License Attribution 4.0 International (CC BY 4.0).
have highlighted opportunities both for applying CBR to explain other systems
and for explaining the behavior of CBR systems.
This paper focuses on needs for explaining CBR system behavior. Researchers
have observed that simply presenting the most similar case in a case base may not
be suﬃcient for optimal explanation (e.g., [11, 31, 32, 43]). A challenge for CBR
is assessing the explanations that are generated. The obvious approach—and the
only deﬁnitive method—for assessing the usefulness of explanations to end users
is human subjects studies. However, such studies require substantial resources
and few have been done for CBR explanation systems. Keane and Kenny refer
to this gap as “[t]he embarrassment of user testing.” In their extensive survey
of twinned CBR-ANN systems, they “found less than a handful (i.e., < 5) that
performed any adequate user testing of the proposal that cases improved the
interpretability of models” [20]. Belén Dı́az Agudo’s ICCBR-19 keynote talk on
explanation listed human subjects studies of human-oriented explanations as one
of four key explanation challenges for the CBR community.1
This paper aims to solidify understanding of the state of the art in assessing
explanation of CBR systems. First, it presents general background on the nature
of explanation and human subjects evaluation in XAI. Second, it presents a
survey of human-subjects evaluations of explanation capabilities in CBR, to
provide a reference for the methods and capabilities that have been assessed.
Third, it categorizes this work, to identify the space to be explored and current
gaps. Finally, based on the categorization and observations of prior research, it
proposes ﬁve steps—both short and longer term—for advancing the evaluation
of explanation of CBR systems and their explanation capabilities.

2 Aims of Explanation

Numerous studies have examined what makes a good explanation and how to
explain in XCBR. Kenny et al. [21] propose a “tricorn” user model for XAI,
with explanations relating to a user’s domain model, model of the system, and
model of how particular explanation strategies (e.g., graphical presentation of
information) relate evidence to a conclusion. Schoenborn et al. [45] survey ex-
plainable CBR and Sormo, Cassens and Aamodt [47] survey explanation in CBR
with an emphasis on criteria for good explanations. In general, explanations can
serve a rich range of needs, depending on the user’s knowledge state, goals, and
tasks [24]. Here we focus on two fundamental goals. One is transparency, under-
standing how the system produced its answer. This may be useful both to foster
trust (e.g., [14]) and to diagnose problems to guide system repair or reﬁnement
of system knowledge or processes. The other is justiﬁcation, understanding why
the system’s answer is a good answer [47].

1
https://iccbr2019.com/wp-content/uploads/2019/09/Keynote-BelenDiazAgudo.pdf
3 Assessing Explanation Systems
Given the goal of serving a human user, the gold standard for system assess-
ments must be the results of human subjects studies. Human subjects studies
can be designed directly to assess and compare explanation quality of alterna-
tive systems (e.g., [10, 30, 37, 38]). However, such assessments are challenging to
design and may be time-consuming and expensive, especially given that many
aspects of system processing (e.g., situation assessment, similarity assessment,
or adaptation) might need to be explained. As a result, few XCBR eﬀorts use
human subject evaluations and the space of possible factors aﬀecting explana-
tion eﬀectiveness is sparsely covered. Some studies rely on informal evaluations,
seeking reactions of a small set of subjects, and many system designs simply rely
on the system-builder’s intuitions. A goal of this paper is to provide a foundation
for choosing aspects of a system to evaluate and to help identify the aspects that
are most and least understood, to illuminate areas for future study.

4 Categorizing Explanations of CBR Systems
This section presents a categorization of research on explaining CBR systems.
We begin by introducing the dimensions used and then place research accord-
ing to those dimensions. The dimensions primarily consider the aspects of the
CBR process being explained and the knowledge (derived from the CBR knowl-
edge containers) that is the focus of those explanations. Of particular interest is
identifying systems that have been evaluated with human subjects studies and
opportunities for future human subjects studies. Schoenborn et al. [45] present
a complementary survey, categorizing XCBR systems according to four types
of aspects: deﬁnitional (explanation goal and kinds such as why-explanations or
how-explanations), model-based or model-agnostic explanation, and (overlap-
ping with this survey) the medium of presentation. We note that XCBR is not
limited to explanation of CBR systems, including, for example, much interesting
work on CBR to support explanation in non-CBR systems (such as [20, 41]).
However, that work is outside the scope of this paper.
Explanations of the conclusions of a CBR system may focus on the solution
provided by the system or the results of any of the intermediate steps in the
CBR cycle—retrieve, reuse, revise, or retain [1]—with an account of how those
results contribute to the solution. For example, CBR system behavior could be
explained by the retrieved case, or a retrieved case plus how it was adapted. In
principle, the retrieval could be explained both by similarity criteria and by how
previous retention decisions aﬀected the choice of cases to retain.
For any of the steps, the generation of a particular result depends on the
knowledge brought to bear. Knowledge from any of the CBR knowledge contain-
ers [42] could be relevant to an explanation. In addition, given the well known
tight coupling of knowledge containers, there may be a choice in determining
which type of knowledge to highlight for a user (e.g., indexing or similarity
knowledge accounting for retrieval of a case, if case relevance may not be obvi-
ous to the user, or case adaptation knowledge, if case relevance is intuitive but
the adapted solution may not be). Each of these choices provides a dimension
for categorizing explanations of CBR systems.

4.1 Categorization Table
The table in Figure 1 categorizes the explanation approaches taken by sixteen
projects. Research evaluated with human subjects studies is highlighted in bold.
Rows are divided into four sections. The ﬁrst section categorizes the research by
Explanation Medium, the presentation of the explanation. The medium may be
in graphical/visual or textual (explanations in internal system symbolic repre-
sentations are considered to be in textual form). The second section categorizes
the research by the Target of Explanation. Targets of the explanations may be
the system’s non-adapted ﬁnal result, the reason for the retrieval of the case
used, the result when the retrieved case was adapted, or aspects of the ongoing
system process.
The third section of rows categorizes the research by the primary Knowledge
Focus of the explanation. A holistic focus provides an entire retrieved case to
the user; the user determines how the case relates to explanatory needs. A key
features focus provides information about the case features salient for retrieval.
An index knowledge focus explains retrievals in terms of memory organization,
the knowledge used to generate indices from cases, and/or knowledge underlying
how those indices may be revised or transformed during the retrieval process.
An adaptation knowledge focus explains the generation of a solution in terms of
the application of knowledge such as adaptation rules, adaptation cases, or other
knowledge from the adaptation knowledge container. A similarity focus explains
retrieval in terms of the similarity measure and/or analysis of case features in
terms of similarity criteria. In principle, a system explaining CBR system behav-
ior could have multiple knowledge focuses. However, in practice, most research
has focused on single focus types. The ﬁnal two sections of rows categorize the
research by Explanation Recipient and the Process or Result Focus. The recipient
of the explanation may be a human or an AI system (as in systems that explain
internally to guide processing). The focus of explanation is the ﬁnal the system
process used to obtain it (for transparency) or system result (for justiﬁcation).

4.2 The Systems Categorized
Armengol, Ontanón and Plaza [4] present a method for explaining similarity
to justify solutions in classiﬁcation tasks. The approach was applied to provide
justiﬁcation to other agents in a multi-agent system. This system focuses on ex-
plaining the similarity of the non-adapted ﬁnal result in text (symbolic) format,
for AI systems.
Burke and Kass [7] present a system that explains the retrieval result for
video clips with “bridging text” accounting for the relevant indices, focusing on
index knowledge, for presentation to a human user.
Cunningham, Doyle and Loughrey [10] performed, to our knowledge, the
ﬁrst human subjects study on prior cases as explanations. Their study, using
Cunningham, Doyle & Loughrey, 03

Nugent, Cunningham & Doyle, 05
Doyle, Cunningham & Walsh, 06

Maximini, Freßman & Schaaf, 04
Armengol, Ontanón & Plaza, 04

Massie, Craw & Wiratunga, 04

Muhammad et al., 15
Burke & Kass, 96

Lieber et al, 08
Lamy et al, 19

McSherry, 04
McSherry, 01

Ong et al., 97
Kolodner, 83

Leake, 91
Kass, 90
a
Explanation Visualization  
Medium Text              
Non-Adapted
         
Final Result
Retrieval
  
Target for Result
Explanation Adaptation
 
Result
Ongoing

Process
Case Only
     
(Holistic)
Key Features    
Knowledge Index
  
Focus Knowledge
Adaptation
 
Knowledge
Similarity  
Explanation Human User              
Recipient AI System  
Process or System Process     
Result Focus System Result           

Fig. 1. A Categorization of Explanations of Case-Based Reasoning.

37 subjects, compared the convincingness of no explanation, presentation of the
case on which a CBR system’s conclusion was based, and rule-based explanation
for predictions of blood alcohol levels. The system’s target is the non-adapted
ﬁnal result in text format, with holistic knowledge focus, to a human user.
Doyle, Cunningham and Walsh [12] study explanation via a selected prior
case (not necessarily the nearest case), justiﬁcation text and a conﬁdence level,
in the domain of hospital admission and discharge decisions. The justiﬁcation
text highlighted features that support or undermine relevance of the explanation
case. The study had two evaluation stages, a formative evaluation and a second
stage of two studies, one with 106 domain experts judging conﬁdence in the
system and the appropriateness of the explanation case, and the second with 14
domain experts judging whether the nearest neighbor or explanation case better
explained the system decision. The system’s target is the non-adapted ﬁnal result
in text format, with holistic knowledge focus, to a human user.
Kass [19] presents a case adaptation system that provides textual descriptions
of adaptation paths that could be presented to explain adaptations to a user.
The system’s target is the adaptation result in text format, with knowledge focus
of adaptation knowledge, for a human user.
Kolodner [22] presents explanation of retrieval by a question transformation
trace leading to the retrieval result in a text format, with the knowledge focus
index knowledge to a human user.
Lamy et al. [23] focus on visualization of diﬀerences and similarities between
a query case and similar cases, in the context of a CBR system to diagnose breast
cancer. Their system was evaluated by 11 users who selected a treatment based
on the data presented by the system and indicated their conﬁdence level in the
choice. The target is the non-adapted ﬁnal result in the form of a visualization,
with a holistic knowledge focus, to a human user.
Leake [24] presents a set of criteria for evaluating CBR-generated explana-
tions of anomalous events in stories for various goals (e.g., prediction of an event
or prevention of an event). The system’s target is the non-adapted ﬁnal result in
text format, with a knowledge focus of holistic, for an AI system.
Lieber et al. [29] present an approach to explaining adaptations, in the con-
text of breast cancer treatment, by presenting adaptation paths with annotations
of each step. The system’s target is the adaptation result in text format, with
knowledge focus on adaptation knowledge, to a human user.
Massie, Craw and Wiratunga [30] present a visualization tool aimed at pre-
senting feature similarity between a new problem and retrieved case, in the
domain of drug design. The work is assessed by asking two domain experts to
assess similarity and express their conﬁdence in the answer the solution provided.
The target for explanation is the retrieval result in the form of a visualization,
with a holistic knowledge focus and explanation to a human user.
Maximini, Freßman, and Schaaf [31] present a system including a range of
explanation capabilities for an intellectual property domain. The system includes
two explanation components. The ﬁrst explains the set of cases retrieved in
terms of attributes and supports user navigation through the candidates, as well
as presenting the most decisive attributes to narrow the retrieval set. When
no candidates are retrieved, it can also help the user understand which query
features could not be matched, to aid query reﬁnement. The other supports
examination of the result, with color coding of features by similarity as well as
a graphical visualization and textual explanation of feature connections. The
system’s target is the non-adapted ﬁnal result in text format, with a knowledge
focus of the similarity for a human user.
McSherry [32] presents an explanation method that provides the features of
the nearest cases that support or oppose a prediction. This system focuses on
explaining the non-adapted ﬁnal result in form of case text, with knowledge focus
on key features to a human user.
McSherry [33] presents an explanation method for conversational CBR for
diagnosis of computer faults. The system provides explanations of the relevance
of questions. This system focuses on explaining the ongoing process in the form
of text, with knowledge focus on index generation, to a human user.
Muhammad et al. [34] present an approach to explaining recommendations
based on positive and negative features compared to user preferences, in a hotel
recommendation domain. The recommender personalizes explanations by high-
lighting the features the given user seeks. This system focuses on explaining the
non-adapted ﬁnal result in form of text, with a knowledge focus on key features
to a human user.
Nugent, Cunningham and Doyle [37] extend their original study [10] for an ex-
planation component providing the retrieved case, text explaining the diﬀerence
between the query and explanation case, and a conﬁdence value. If conﬁdence is
below a threshold, an additional counter case is provided, with a description of
how feature values in both cases impact classiﬁcation. The study used 12 human
subjects. The system target is the non-adapted ﬁnal result in text format, with
a holistic knowledge focus and explanation for a human user.
Ong et al. [39] present a system that explains similarity by presenting sim-
ilarity scores for matching features, in a cancer diagnosis domain. This system
focuses on explaining the non-adapted ﬁnal result in form of text, with knowledge
focus on the case as a whole, so holistic, to a human user.

5 Next Steps

Examination of research in the survey suggests ﬁve steps for advancing evalua-
tion of explanation of CBR systems. The ﬁrst, building on the dimensions of the
categorization, is to ﬁll gaps in coverage of evaluation dimensions in the cate-
gorizations. The second and third involve community eﬀort to make evaluations
more practical to perform and results more comparable. This requires both ad-
vancing automated evaluations, to enable gathering initial evaluation data more
rapidly, and establishing standard datasets and standardized evaluation designs,
in order to make studies more useful for understanding the relationships between
diﬀerent explanation focuses and how they contribute to the beneﬁt of explana-
tion. The fourth and ﬁfth are to move beyond the current space of explanations,
focused on local explanation of a single solution, to build explanations to help
users better understand how systems will perform in the future and to explain
system failures. Understanding the full scope of explanation of CBR systems will
require experiments that assess the full system, including eﬀects of component
interactions.

5.1 Using the Categorization Scheme to Suggest Evaluation Targets

The categorization table of the previous section shows the distribution of hu-
man subjects studies for explaining CBR. We observe that existing evaluations
have primarily focused on case-oriented explanations—explaining by presenting
cases and highlighted features to the user. This has been important for assessing
the fundamental CBR premise that cases are a natural mode of explanation.
However, in domains unfamiliar to the user, explanatory support for assessing
similarity may be crucial. This received some recognition [4, 30], but so far, to
our knowledge, only one method has been tested with human subjects, and in
a small-scale assessment (by two experts) [30]. Adaptation-focused explanation
has also received little attention.
The survey also suggests the potential for studies of how diﬀerent modes of
information presentation aﬀect eﬃcacy of explanations. Systematic evaluation
will require understanding:

– Alternative methods and tradeoﬀs for conveying to the user information
about system knowledge in each of the CBR knowledge containers [42]: the
vocabulary, similarity measure, case base, and adaptation knowledge.
– How best to explain the processes underlying each step of the CBR cycle:
retrieve, reuse, revise, and retain [1].
– The inﬂuence of information needs [24, 47]: how to select parts of that process
to highlight to particular users.

The ideal result would be a toolkit of best practices for explaining classes of CBR
systems based on the tasks they perform. For example, explaining a “compare
and contrast” task, as in American legal reasoning [5], will necessarily diﬀer from
explaining anomalies [24]. Expert systems research on generic tasks [8] suggests
the potential to identify a comparatively limited set of task classes. This in
turn would increase the practicality and usefulness of human subjects studies by
encouraging studies with wide task applicability. A community eﬀort to develop
such a collection could leverage both XCBR research and the application of
XCBR in domains for which explanation is important.

5.2 Advancing Automatic Evaluation

There is a longstanding tradition of testing conversational CBR systems with
simulated users (e.g., [3]). Research on case-based explanation generation has ex-
amined criteria for evaluating the goodness of explanations generated [18, 24, 46].
This prior work raises an interesting question: Would it be possible to develop
automated evaluation schemes to test aspects of the explanations generated by
CBR systems? Even if such methods did not have complete ﬁdelity to human
subjects studies, automatic testing could be a valuable intermediate step.
Work on case-based explanation for story understanding systems has devel-
oped criteria for evaluating explanations according to their ability to resolve
anomalies in stories, address system questions, and satisfy other explanation
goals [24, 40]. Applying such methods to a broader class of tasks would require
analysis of the information needs associated with those tasks and how they can
be satisﬁed. Providing a general solution for this would be an enormous task.
However, it could be more feasible in testbed microdomains—were the commu-
nity to embrace developing such resources.
5.3 Developing Standardized Evaluation Resources
The ability to compare and replicate human subjects studies is impeded by the
lack of agreed-upon testbeds for such evaluations. Diﬀerent research groups use
diﬀerent data sets, making it more diﬃcult to assess the relative contributions
of explanations of diﬀerent parts of the CBR process and how those components
interact. From the ﬁve human subject studies reported in this paper’s categoriza-
tion, two (done by the same research group) used the blood-alcohol domain [10,
37], one used the tablet formulation domain [30], one used Bronchiolitis medical
data [12], and one used breast cancer data [23]. Of the 4 unique domains, only
the breast cancer data [13] is publicly available. None of these data sets involve
CBR systems including case adaptation.
The development of the needed data sets—and agreement to use them—must
be a community endeavor. The CBR community has a history of developing
shared resources and advancing the state of the art with competitions, as in
the Computer Cooking Competition series (e.g., [35]). This series included var-
ious forms of human evaluations, both by human judges and by the conference
attendees as a whole. An explanation challenge, based on standard tasks and
data sets and informally evaluated by humans as part of the competition, could
“jump start” the development of standard resources. In addition, the design of
the competition could help establish evaluation protocols to form the basis for
more formal human subjects studies.

5.4 Explaining the prior cases themselves
Explanations based on presenting prior cases assume the quality of the retrieved
case can be trusted—that no justiﬁcation is needed for why the case should
be believed. However, especially if cases have been automatically generated by
adaptation processes, understanding and accepting a case as an explanation
depends on understanding the presented case and its relevance. That may depend
on understanding how it was derived—its provenance [28]. The use of provenance
for explaining CBR is an interesting area for future research.

5.5 Broadening CBR Explanations to Explaining Failures
Existing studies of explaining CBR focus primarily on accounting for why the
system reasoned as it did. Users may use that information to determine the
applicability of a proposed solution. This is suﬃcient to determine trust in the
current solution. However, when users determine that the system went astray, the
explanation components generally have no capability to explain why the CBR
process failed. In the absence of such an explanation it is harder to determine
trust in the system’s future processing.
Some CBR research has applied metareasoning to explain and repair reason-
ing failures within the CBR process for internal system learning (e.g., [9, 15]).
This points to another topic for CBR system explanation: accounting for expec-
tation failures in how the system operates, to help users reﬁne their predictive
model of the CBR system overall.
6 Conclusion
This paper has presented an initial categorization of research on explaining CBR
systems, highlighting dimensions for which human subjects evaluations have
been done—and for which they are still needed. Based on this categorization
it has proposed XCBR research opportunities including increased attention to
explaining similarity and adaptation. It has also advocated new initiatives on
automatic evaluation, building community evaluation resources, explaining case
provenance, and explaining system failures.

7 Acknowledgments
We acknowledge support from the Department of the Navy, Oﬃce of Naval Re-
search (Award N00014-19-1-2655), and the US Department of Defense (Contract
W52P1J2093009).

References
1. Aamodt, A., Plaza, E.: Case-based reasoning: Foundational issues, methodological
variations, and system approaches. AI Communications 7(1), 39–52 (1994)
2. Aha, D., Agudo, B.D., Garcia, J.R. (eds.): Proceedings of XCBR-2018: First work-
shop on case-based reasoning for the explanation of intelligent systems. ICCBR
(2018), http://iccbr18.com/wp-content/uploads/ICCBR-2018-V3.pdf
3. Aha, D., Breslow, L.: Reﬁning conversational case libraries. In: Proceedings of the
Second International Conference on Case-Based Reasoning. pp. 267–278. Springer
Verlag, Berlin (1997)
4. Armengol, E., Ontanón, S., Plaza, E.: Explaining similarity in CBR. In: ECCBR
2004 Workshop Proceedings. pp. 155–164 (2004)
5. Ashley, K., Rissland, E.: Compare and contrast, a test of expertise. In: Proceedings
of the Sixth Annual National Conference on Artiﬁcial Intelligence. pp. 273–284.
AAAI, Morgan Kaufmann, San Mateo, CA (1987)
6. B. Diaz Agudo, J.R.G., Watson, I. (eds.): Proceedings of XCBR-2019: Second
workshop on case-based reasoning for the explanation of intelligent systems. IC-
CBR (2018)
7. Burke, R., Kass, A.: Retrieving stories for case-based teaching. In: Leake, D. (ed.)
Case-Based Reasoning: Experiences, Lessons, and Future Directions, pp. 93–109.
AAAI Press, Menlo Park, CA (1996)
8. Chandrasekaran, B.: Generic tasks in knowledge-based reasoning: High level build-
ing blocks for expert system design. IEEE Expert 1(3), 23–30 (1986)
9. Cox, M., Ram, A.: Introspective multistrategy learning: On the construction of
learning strategies. Artiﬁcial Intelligence 112(1-2), 1–55 (1999)
10. Cunningham, P., Doyle, D., Loughrey, J.: An evaluation of the usefulness of case-
based explanation. In: Case-Based Reasoning Research and Development: Proceed-
ings of the Fifth International Conference on Case-Based Reasoning, ICCBR-03.
pp. 122–130. Springer-Verlag, Berlin (2003)
11. Doyle, D., Cunningham, P., Bridge, D., Rahman, Y.: Explanation oriented re-
trieval. In: Funk, P., González Calero, P.A. (eds.) Advances in Case-Based Rea-
soning. pp. 157–168. Springer, Berlin, Heidelberg (2004)
12. Doyle, D., Cunningham, P., Walsh, P.: An evaluation of the usefulness of explana-
tion in a case-based reasoning system for decision support in bronchiolitis treat-
ment. Computational Intelligence 22(3-4), 269–281 (2006)
13. Dua, D., Graﬀ, C.: UCI machine learning repository (2017),
http://archive.ics.uci.edu/ml
14. Floyd, M., Aha, D.: Special issue on case-based reasoning. AI Communications
30(3-4), 281–294 (2017)
15. Fox, S., Leake, D.: Introspective reasoning for index reﬁnement in case-based rea-
soning. The Journal of Experimental and Theoretical Artiﬁcial Intelligence 13(1),
63–88 (2001)
16. Goodman, B., Flaxman, S.: European Union regulations on algorithmic decision-
making and a “right to explanation”. AI Magazine 38(3), 50–57 (2017)
17. Gunning, D., Aha, D.W.: DARPA’s explainable artiﬁcial intelligence program. AI
Magazine 40(2), 44–58 (2019)
18. Johs, A.J., Lutts, M., Weber, R.O.: Measuring explanation quality in XCBR. In:
Minor, M. (ed.) Workshop proceedings ICCBR 2018. pp. 75–83 (2018)
19. Kass, A.: Developing Creative Hypotheses by Adapting Explanations. Ph.D. thesis,
Yale University (1990), northwestern University Institute for the Learning Sciences,
Technical Report 6
20. Keane, M.T., Kenny, E.M.: How case-based reasoning explains neural networks: A
theoretical analysis of XAI using post-hoc explanation-by-example from a survey
of ANN-CBR twin-systems. In: Case-Based Reasoning Research and Development:
ICCBR-19. pp. 155–171. Springer, Berlin (2019)
21. Kenny, E.M., Ford, C., Quinn, M., Keane, M.T.: Explaining black-box classiﬁers
using post-hoc explanations-by-example: The eﬀect of explanations and error-rates
in XAI user studies. Artiﬁcial Intelligence 294, 103459 (2021)
22. Kolodner, J.L.: Reconstructive memory: A computer model. Cognitive Science
7(4), 281–328 (1983)
23. Lamy, J.B., Sekar, B., Guezennec, G., Bouaud, J., Séroussi, B.: Explainable artiﬁ-
cial intelligence for breast cancer: A visual case-based reasoning approach. Artiﬁcial
intelligence in medicine 94, 42–53 (2019)
24. Leake, D.: Goal-based explanation evaluation. Cognitive Science 15(4), 509–545
(1991)
25. Leake, D.: CBR in context: The present and future. In: Leake, D. (ed.) Case-Based
Reasoning: Experiences, Lessons, and Future Directions, pp. 3–30. AAAI Press,
Menlo Park, CA (1996), http://www.cs.indiana.edu/˜leake/papers/a-96-01.html
26. Leake, D.: Cognition as case-based reasoning. In: Bechtel, W., Graham, G. (eds.)
A Companion to Cognitive Science, pp. 465–476. Blackwell, Oxford (1998)
27. Leake, D., McSherry, D.: Introduction to the special issue on explanation in case-
based reasoning. Artiﬁcial Intelligence Review 24(2), 103–108 (2005)
28. Leake, D., Whitehead, M.: Case provenance: The value of remembering case
sources. In: Case-Based Reasoning Research and Development: ICCBR-07. pp.
194–208. Springer-Verlag, Berlin (2007)
29. Lieber, J., d’Aquin, M., Badra, A., Napoli, A.: Modeling adaptation of breast
cancer treatment decision protocols in the KASIMIR project. Applied Intelligence
28(3), 261–274 (2008)
30. Massie, S., Craw, S., Wiratunga, N.: A visualisation tool to explain case-base rea-
soning solutions for tablet formulation. In: Proceedings of the 24th SGAI Interna-
tional Conference on Innovative Techniques and Applications of Artiﬁcial Intelli-
gence. Springer-Verlag, Berlin (2004)
31. Maximini, R., Freßmann, A., Schaaf, M.: Explanation service for complex CBR ap-
plications. In: Advances in Case-Based Reasoning, ECCBR. pp. 302–316. Springer,
Berlin (2004)
32. McSherry, D.: Explaining the pros and cons of conclusions in CBR. In: Advances
in Case-Based Reasoning. pp. 317–330. Springer, Berlin (2004)
33. McSherry, D.: Interactive case-based reasoning in sequential diagnosis. Applied
Intelligence 14(1), 65–76 (2001)
34. Muhammad, K., Lawlor, A., Rafter, R., Smyth, B.: Great explanations: Opinion-
ated explanations for recommendations. In: Case-Based Reasoning Research and
Development. pp. 244–258. Springer, Berlin (2015)
35. Najjar, N., Wilson, D.: Computer cooking contest workshop at the twenty-
ﬁfth international conference on case-based reasoning (ICCBR 2017). In: ICCBR
2017 Workshops, Doctoral Consortium, and Competitions (2017), http://ceur-
ws.org/Vol-2028/XXCCC17 preface.pdf
36. Nugent, C., Cunningham, P.: A case-based recommender for black-box systems.
Artiﬁcial Intelligence Review 24(2), 163–178 (2005)
37. Nugent, C., Cunningham, P., Doyle, D.: The best way to instil conﬁdence is by
being right. In: International Conference on Case-Based Reasoning. pp. 368–381.
Springer (2005)
38. Nugent, C., Doyle, D., Cunningham, P.: Gaining insight through case-based expla-
nation. Journal of Intelligent Information Systems 32(3), 267–295 (Jun 2009)
39. Ong, L.S., Shepherd, B., Tong, L.C., Seow-Choen, F., Ho, Y.H., Tang, C.L., Ho,
Y.S., Tan, K.: The colorectal cancer recurrence support (CARES) system. Artiﬁcial
Intelligence in Medicine 11(3), 175–188 (1997)
40. Ram, A.: AQUA: Asking questions and understanding answers. In: Proceedings
of the Sixth Annual National Conference on Artiﬁcial Intelligence. pp. 312–316.
Morgan Kaufmann, Seattle, WA (July 1987)
41. Recio-Garcı́a, J.A., Dı́az-Agudo, B., Pino-Castilla, V.: CBR-LIME: A case-based
reasoning approach to provide speciﬁc local interpretable model-agnostic expla-
nations. In: Case-Based Reasoning Research and Development: ICCBR-20. pp.
179–194. Springer, Online (2020)
42. Richter, M.: Introduction. In: Lenz, M., Bartsch-Spörl, B., Burkhard, H.D., Wess,
S. (eds.) CBR Technology: From Foundations to Applications, chap. 1, pp. 1–15.
Springer, Berlin (1998)
43. Roth-Berghofer, T.: Explanations and case-based reasoning: Foundational issues.
In: Advances in Case-Based Reasoning. pp. 389–403. Springer, Berlin (2004)
44. Rudin, C.: Stop explaining black box machine learning models for high stakes
decisions and use interpretable models instead. Nature Machine Intelligence 1,
206–215 (2019)
45. Schoenborn, J., Weber, R., Aha, D., Cassens, J., Althoﬀ, K.: Explainable case-
based reasoning: A survey. In: Proceedings of the AAAI Explainable Agency in
Artiﬁcial Intelligence Workshop (2020)
46. Sizov, G., Öztürk, P., Bach, K.: Evaluation of explanations extracted from tex-
tual reports. In: Proceedings of the Twenty-Ninth International Florida Artiﬁcial
Intelligence Research Society Conference, FLAIRS-16. pp. 425–429. AAAI Press
(2016)
47. Sormo, F., Cassens, J., Aamodt, A.: Explanation in case-based reasoning—
perspectives and goals. Artiﬁcial Intelligence Review 24(2), 109–143 (2005)