=Paper=
{{Paper
|id=Vol-3769/paper11
|storemode=property
|title=Arguments Based on Domain Rules in Prediction Justifications
|pdfUrl=https://ceur-ws.org/Vol-3769/paper11.pdf
|volume=Vol-3769
|authors=Joeri Peters,Floris Bex,Henry Prakken
|dblpUrl=https://dblp.org/rec/conf/cmna/PetersBP24
}}
==Arguments Based on Domain Rules in Prediction Justifications==
https://ceur-ws.org/Vol-3769/paper11.pdf
Arguments Based on Domain Rules in Prediction
Justifications
Joeri Peters1,2,∗ , Floris Bex1,3 and Henry Prakken1
1 Utrecht University, Utrecht, The Netherlands
2 Netherlands National Police, Driebergen, The Netherlands
3 Tilburg University, Tilburg, The Netherlands
Abstract
Ensuring the interpretability of trained machine learning models is often paramount, particularly in high-stakes
domains such as counter-terrorism and other forms of law enforcement. Post hoc techniques have emerged as a
promising avenue for justifying the predictions of complex models. However, while these approaches provide
valuable insights, they often lack the ability to directly reference familiar domain rules and make use of these
rules to guide explanations. This paper introduces a method for incorporating arguments about the applicability of
domain rules in justifying classifier predictions.
Keywords
Case-Based Argumentation, Precedential Constraint, Explainable AI, Domain Knowledge
1. Introduction
This paper is concerned with explainability in machine learning (ML). Specifically, we focus on enhancing
the explainable artificial intelligence (XAI [12]) approach known as ‘a fortiori case-based argumentation’
(AF-CBA [17]). AF-CBA justifies binary classification predictions using the theory of precedential
constraint [10], that is, referencing precedential cases from a case base constructed from training (or
historical [15]) data. Our goal is to extend this framework by incorporating domain rules, recognising
that domain-specific knowledge plays a pivotal role in decision-making processes.
ML models are often regarded as ‘black boxes’ when their opacity is high, whether due to relative
complexity or proprietary protection [11, 8]. Neural networks serve as a typical example of intricate
models that have revolutionised predictive accuracy at the cost of increased opacity. Transparency and
explainability concerns become particularly critical in high-stakes domains, such as law enforcement,
where decisions may carry significant consequences for individuals or court cases. Predictions have to
be highly accurate—possibly necessitating opaque models—yet explainable. Post hoc approaches like
AF-CBA are aimed at solving this problem by justifying ML predictions ‘after the fact’, meaning that the
approach does not access the ML model itself and is therefore model agnostic. In our experience, the
need for such an approach arises relatively frequently in practice. ML models can be inaccessible at the
moment an explanation is required or the type of explanation it can offer is too technical for the intended
users, thereby rendering it a black box. Furthermore, there can be situations when the performance
metrics of an interpretable alternative to black box approaches is deemed unsatisfactory, necessitating a
post hoc solution. AF-CBA produces such justifications on the basis of earlier cases (precedents).
Applicable scenarios can be drawn from the domain of counter-terrorism, where ML classifiers can be
used to quickly yet objectively distinguish between two outcomes. For example, there may be a need
to decide whether a particular incident is the responsibility of a specific terrorist organisation, judging
by the modus operandi and objectives of its members. Another binary categorisation could be between
the incident forming a part of a large-scale coordinated attack and it being a ‘lone-wolf’ incident, the
CMNA’24: The 24th International Workshop on Computational Models of Natural Argument, September 17, 2024, Hagen,
Germany
∗ Corresponding author.
$ j.g.t.peters@uu.nl (J. Peters); f.j.bex@uu.nl (F. Bex); h.prakken@uu.nl (H. Prakken)
0009-0009-1493-9872 (J. Peters); 0000-0002-5699-9656 (F. Bex); 0000-0002-3431-7757 (H. Prakken)
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
outcome of which warrants a different police response. As a running example, we adopt the scenario
in which officials seek to determine whether a violent event should be classified as an act of terrorism.
It is realistic that a classifier should be used in order to facilitate quick yet valid judgement in such a
situation, to avoid responders acting on gut feeling alone. However, the number of applicable precedents
is relatively low and much contextual knowledge is involved in making this decision. Hence, the system
should be transparently constrained by experts’ knowledge of this domain. Our approach is in line with
a tradition of viewing rule- and case-based reasoning as complementary. For instance, the two were
combined in an overall architecture by Golding & Rosenbloom [7] to allow the latter to produce analogies
in order to handle exceptions to the (incomplete) rule set. A similar integration of rules and cases was
used by Rissland & Skalak [18] in their CABARET system, aimed at an area of income tax law. Our
goal, however, is to let AF-CBA make use of and refer to such domain knowledge in its justifications of
the predicted outcomes of a black-box model.
The rest of this paper is structured as follows. We introduce our XAI approach in Section 2 before
considering how to incorporate domain rules in Section 3. Finally, we discuss conclusions and future
work directions in Section 4.
2. Preliminaries
In justifying binary class labels, the predictions of a classifier trained on labelled data during its training
phase can be likened to court decisions based on judicial precedents. In this vein, Prakken & Ratsma [17]
propose a top-level model, afterwards dubbed AF-CBA, drawing on AI & Law research and utilising
case-based argumentation inspired by Horty’s model of a fortiori reasoning [9]. AF-CBA is influenced
by CATO [1] and work by Čyras et al. [4, 3, 5]. Contrary to [3], AF-CBA is not its own explainable
classification approach, but a post hoc approach used to justify the classification predictions of another
ML model.
Construct Population Label
case base (unlabelled) 0..n data
1
Case Labelled
Focus
base dataset
case
Justify Train
Classifier
prediction classifier
(black box)
Prediction
Justification
Figure 1: A schematic depiction of AF-CBA’s workflow. The case base is constructed either instantly on the
basis of the labelled data (dashed line) or stepwise on the basis of earlier predictions (dotted line, as in [15]).
The context of AF-CBA is illustrated in Figure 1. A labelled dataset constitutes a random sample from
the overall population, to which annotators or decision-makers assign labels, and on which a classifier is
trained (supervised ML). A focus case represents a single, random sample from the same population, and
the classifier assigns a predicted outcome to it. Due to the black-box nature of the classifier, it lacks the
capability to explain the rationale behind the prediction. AF-CBA addresses this limitation by utilising
either the labelled set or an archive of previous case predictions [15] as a case base, engaging in an
argument game between a proponent and an opponent of the predicted outcome. In this argument game,
cases which are similar to the focus case are cited in order to argue that the focus case ought to receive
the same outcome. The decision is forced when no relevant differences exist between the focus case and
the precedent. Moves in the argumentation game follow Dung’s abstract argumentation framework [6],
with the game modelled on grounded semantics [16]. The notion of precedential constraint is that a focus
case ought to receive the same outcome as a precedential case if any differences between those cases only
serve to strengthen the focus case for that particular outcome. A winning strategy for the proponent is
then presented as a justification for the predicted outcome in the form of an argument graph.
An abstract argument framework (AF), introduced by Dung [6], consists of a pair AF = ⟨A, attack⟩,
where A represents a set of arguments, and attack is a binary relation on A. A subset B of A is termed
conflict-free if no arguments in B attacks arguments in B and admissible if it is both conflict-free and
capable of defending itself against attacks. In other words, if an argument A1 is in B, and some argument
A2 in A attacks A1 , then some argument in B must attack A2 . There are different types of admissible sets,
known as extensions. We focus on the grounded extension, which has the additional properties that it
contains all arguments it defends and is subset-minimal for these conditions.
Formally, a case in the case base (CB) comprises an outcome and a fact situation. The case’s outcome
is a binary label, denoted as o or o′ . Variables s and s̄ represent the two sides, such that s = o if s̄ = o′
and vice versa. The fact situation includes dimensions (features), where each dimension is a tuple
d = (V, ≤o , ≤o′ ). The tuple consists of a value set V and two partial orderings on V , ≤o and ≤o′ , such
that v ≤o v′ if and only if v′ ≤o′ v for v, v′ ∈ V . Each dimension has a tendency, with a positive tendency
indicating a higher value is associated with one outcome (e.g., 1 or true), and vice versa for the other.
The tendency is sometimes given explicitly, that is: di+ or di− . A value assignment, represented as (d, v),
signifies the value x of dimension d in case c ∈ CB as v(d, c) = x. The collective value assignments for
all dimensions d in the non-empty set D form a fact situation denoted as F. We assume that two fact
situations pertain to the same set D. Defining a case as c = (F, outcome(c)) where outcome(c) ∈ {o, o′ },
the fact situation of case c can be expressed as F(c).
When assessing two fact situations, one may find that one case is ‘stronger’ or ‘better’ for a specific
outcome than the other. The outcome of a focus case is considered forced if there exists a precedent in
the CB with the same outcome, and all differences between the focus case and that precedent serve to
strengthen the focus case for that very outcome [10].
Definition 1 (Preference relation for fact situations). Given two fact situations F and F ′ , F ≤s F ′ iff
v ≤s v′ for all (d, v) ∈ F and (d, v′ ) ∈ F ′ .
Definition 2 (Precedential constraint). Given case base CB and fact situation F, deciding F for s is
forced iff CB contains a case c = (F ′ , s) such that F ′ ≤s F.
A fact situation could be forced for both outcomes o and o′ by different precedents, in which case we
can speak of an inconsistent CB:
Definition 3 (Case base consistency). A case base CB is consistent iff it does not contain two cases
c = (F, s) and c′ = (F ′ , s̄) such that F ≤s F ′ . Otherwise it is inconsistent.
A best precedent has the same outcome as the focus case and as few as possible relevant differences.
Multiple cases can meet these criteria.
Definition 4 (Differences between cases). Let c = (F(c), outcome(c)) and
f = (F( f ), outcome( f )) be two cases. The set D(c, f ) of differences between c and f is D(c, f ) =
{(d, v) ∈ F(c) | v(d, c) ≰outcome( f ) v(d, f )}.
Definition 5 (Best precedent). Let c = (F(c), outcome(c)) and f = (F( f ), outcome( f )) be two cases,
where c ∈ CB and f ∈/ CB. c is a best precedent for f iff:
• outcome(c) = outcome( f ) and
• there is no c′ ∈ CB such that outcome(c′ ) = outcome(c) and D(c′ , f ) ⊂ D(c, f ).
The two players argue about differences between the focus case and precedents from the CB. The
proponent does so in favour of the focus case’s predicted outcome and the opponent to the contrary.
The proponent starts by citing a best precedent. The opponent aims to respond to the proponent’s
initial citation by either presenting a counterexample or making a distinguishing move Worse(c, x) (the
focus case is inferior to precedent c in dimensions x). A distinguishing move can be countered with a
compensation move Compensates(c, y, x) (dimensions y make up for the shortcomings in dimensions x
compared to precedent c). Finally, there is the transformation move Trans f ormed(c, c′ ) (the citation can
be transformed into a case where D(c, f ) = 0). / The proponent can respond using these moves, then the
opponent can do the same in turn, and this back-and-forth continues until the opponent cannot make any
more moves. Note that y in Compensates(c, y, x) can be the empty set. This is intended to guarantee the
possibility of using a compensation move, ensuring the existence of a winning strategy for the proponent
and thus that of a justification for the focus case’s predicted outcome.
Definition 6 outlines the argumentation framework. The compensation move utilises the set sc,
containing compensation definitions. The specifics and structure of sc were intentionally left open
by Prakken & Ratsma [17]. In the most straightforward scenario, sc serves as a partial ordering on
dimensions, indicating, for example, when a high value for di compensates for a low value for d j .
Essentially, sc imparts specific domain knowledge. In this paper, we employ the set sc to explicitly
introduce domain rules into the framework for use by the compensation move.
Definition 6 (Case-based argumentation framework). Given a case base CB, a focus case f ∈ / CB, and
definitions of compensation sc, an abstract argumentation framework AF is a pair < A , attack >, where:
• A = CB ∪ M,
with M = {Worse(c, x) | c ∈ CB, x ̸= 0/ and x = {(d, v) ∈ F( f ) | v(d, f ) outcome( f ) d(c)}
p: Dw ≺ Db
———————————
Conc: compensates(c, Db , Dw )
Table 1
Precedent case c and focus case f .
+ −
Case dcasualties dweapon ... Outcome
c 5 low ... True
f 10 high ... True
In Table 1, f has a higher weapon sophistication (dweapon ) than c, which is associated with non-terrorist
incidents, making f worse than c on this dimension. However, f also has a higher number of casualties
(dcasualties ), which is a strong indicator of a terrorist incident. Using the domain rule that a higher number
of casualties compensates for a higher weapon sophistication, we instantiate the following argument:
COMP( f , c, Db , Dw ):
w: Dw = {dweapon }
b: Db = {dcasualties }
p: {dweapon } ≺ {dcasualties }
———————————
Conc: compensates(c, {dcasualties }, {dweapon })
In this example, the argument states that although f has a higher (‘worse’) weapon sophistication than
c, the higher (‘better’) number of casualties of f compensates for this, justifying the predicted outcome
of true for f on the basis of this precedent.
We assume that the fact situations of both the precedents from the CB and the focus case are known.
Therefore, we cannot argue against the first two premises of this scheme. Furthermore, the scheme is
strict in that the conclusion of this scheme cannot be negated if all its premises are true. But first, we must
consider how we know that premise p (the preference relation underpinning the compensation move) is
true. In practice, conditions may apply for a preference relation and we will now consider the various
forms these conditions may take.
3.2. Conditional Preference Relations
There may be situations where a specific threshold value must be met for a preference relation to be
considered to hold through application of by Scheme 1. For instance, the aforementioned relation
{dweapon } ≺ {dcasualties } may only hold for high numbers of casualties, say at least 4. Fewer casualties
may not be considered a good enough reason to compensate for the fact that the highly sophisticated
weapon used in this incident is so irregular. In other words, the premise p of this instance of Scheme 1
depends on the condition that dcasualties ≥ 4. We consider additional examples of conditions below, but for
now we summarise the sets of conditions for a preference relation Dw ≺ Db with an abstract premise Ψ.
Argumentation Scheme 2 (Preference). Let f be a focus case, s ∈ {o, o′ }, Db , Dw ⊆ D be two sets of
dimensions where Db ∩ Dw = 0, / Ψ be an abstract placeholder whose thruth value represents whether
the preference conditions are fulfilled. Then the preference scheme PREF( f , Db , Dw , D) is defined as the
following reasoning pattern:
ψ: Ψ (preference conditions fulfilled)
===================
Conc: Dw ≺ Db
Scheme 2 evaluates whether Ψ holds in a particular instance. If so, the relevant preference relation
can be concluded and subsequently used as a premise p in the instantiation of Scheme 1. In Table 1, the
focus case f has a ‘worse’ level of sophistication in the weapon used (dweapon ) and a ‘better’ number of
casualties (dcasualties ), with respect to the outcome true. Instantiating Schemes 2 (PREF( f , Db , Dw , D))
and 1 (COMP( f , c, Db , Dw )) lets us construct the following argument:
PREF( f , Db , Dw , D):
ψ: dcasualties ( f ) ≥ 4
===================
Conc: {dweapon } ≺ {dcasualties }
COMP( f , c, Db , Dw ):
w: Dw = {dweapon }
b: Db = {dcasualties }
p: {dcasualties } ≺ {dweapon }
———————————
Conc: compensates(c, {dcasualties }, {dweapon })
Furthermore, there can be more than one threshold as part of Ψ. We could have a preference relation
that states that dcasualties in combination with d f ear (a numerical expression of public fear) is more relevant
than (i.e. is preferred over) dweapon if both dimensions exceed their respective thresholds. We would then
instantiate Scheme 2 as:
PREF( f , Db , Dw , D):
ψ: dcasualties ( f ) ≥ 4 ∧ d f ear ( f ) ≥ 10
===================
Conc: {dweapon } ≺ {dcasualties , d f ear }
Whether certain dimensions surpass certain thresholds is a type of condition that presumes that each
dimension must independently meet a sub-condition. Alternatively, the condition for a preference relation
might hinge on a combination of dimensions, in the form of some evaluation function surpassing a
single threshold. For an example of such a rule, we can imagine a preference relation {dweapon } ≺
{dcasualties , dwounded } and its condition that dcasualties ( f ) + dwounded ( f ) ≥ 10. Here, the evaluation function
is the sum of the number of fatal casualties and non-fatally wounded that compensates for a high level of
weapon sophistication. In other words, the distinction between fatally and non-fatally harmed victims is
of no consequence in this domain rule; what matters is the number of victims.
PREF( f , Db , Dw , D):
ψ: dcasualties ( f ) + dwounded ( f ) ≥ 10
===================
Conc: {dweapon } ≺ {dcasualties , dwounded }
Alternatively, one can imagine rules in which the difference between the number of perpetrators
and victims plays a role in distinguishing terrorist incidents from, say, assassinations. Or perhaps the
ratio between wounded and deceased victims modulates the impact of weapon sophistication in some
hypothetical rule. A weighted mean of several dimensions may have to surpass a certain value. And so
on, domain experts may have dozens of rules for a domain that is particularly well understood and rich in
descriptive dimensions, possibly assisted by some statistical analysis or rule discovery approach. More
complex functions are also possible. One could argue that at least some of such evaluations ought to be
captured in the feature engineering phase before training a model, rather than in post-hoc justifications;
we remind the reader that our approach is model- and data-agnostic, so we should generally support
relevant evaluations.
Consider the following scenario: an attack involving a sophisticated bomb (dweapon ) that does not result
in a high number of casualties (dcasualties ). Under normal circumstances, the sophistication of the weapon
might suggest a targeted assassination rather than a terrorist attack. However, if the event generates
an exceptionally high level of public fear (d f ear ), this could compensate for the lower casualty count,
as the primary goal of terrorism is often to instil fear and disrupt societal normalcy. In this case, the
evaluation function might give significant weight to d f ear , such that a weighted sum of d f ear and dcasualties
is compared to a threshold value.
PREF( f , Db , Dw , D):
ψ: 0.3 · dcasualties ( f ) + 0.7 · d f ear ( f ) ≥ 10
===================
Conc: {dweapon } ≺ {dcasualties , d f ear }
Aforementioned thresholds form conditions on the very dimensions within the preference relation,
Dw and Db . However, there may be situations where contextual factors influence the applicability of the
preference relation. For example, the additional dimension dmeasures (number of security measures in place)
might in certain cases modulate the impact of a the number of casualties in compensating for weapon
sophistication. In this condition, the threshold value pertains to a dimension that is itself not involved in
the preference relation. The conditions for a preference relation can also involve spatiotemporal factors.
For instance, the same set of dimensions might have different thresholds or weights depending on whether
the event occurs in a region currently experiencing political instability. This adaptability is crucial in
counter-terrorism, where the nature of threats and societal impact can change rapidly. When trying
to attribute historical incidents to terrorist organisations, one would have to take into account that an
organisation was founded at a certain moment in time, or was only active within a particular part of the
world. For example, IS (ISIS/ISIL) did not rise to prominence until 2014 in areas of Syria and Iraq. Any
domain rule that is concerned with characteristics of IS incidents or public claims by this organisation
is likely specific to the appropriate time and place. The same type of corncern applies to the Taliban in
Afghanistan before the American invasion in 2001 or their departure in 2021, or the Troubles in Ireland
and Great Britain between 1966 and 1998. For a (simplified) example, consider the following, where the
fact that an incident takes place during the Troubles in Belfast means that {dwounded } compensates for
{dcasualties , dweapon }:
PREF( f , Db , Dw , D):
ψ: dyear ( f ) = 1969 ∧ dlocation ( f ) = Belfast
===================
Conc: {dcasualties , dweapon } ≺ {dwounded }
Alternatively, this particular insight from the domain expert could be used to construct an empty
compensation move on the basis of domain knowledge. Note that if Db = 0, / Scheme 1 describes the
special case of empty compensation. We may want to overlook poor values for {dcasualties , dclaims } out
of hand. Normally in AF-CBA, we allow for compensation moves with Db = 0/ in order to guarantee
a winning strategy (Section 2), as somewhat of an unsatisfactory but necessary default substituting
for a more informative justification. With an argument such as the following, we can actually provide
expert-informed justifications why values in Db are not relevant to the outcome of the focus case despite
not having any compensating dimensions, making an empty compensation move more valuable than it
would otherwise be:
PREF( f , Db , Dw , D):
ψ: dyear ( f ) = 1969 ∧ dlocation ( f ) = Belfast
===================
Conc: {dcasualties , dweapon } ≺ 0/
COMP( f , c, Db , Dw ):
w: Dw = {dcasualties , dweapon }
b: Db = 0/
p: {dcasualties , dweapon } ≺ 0/
———————————
Conc: compensates(c, 0, / {dcasualties , dweapon })
Transitivity (where {d1 } ≺ {d2 } and {d2 } ≺ {d3 } implies {d1 } ≺ {d3 }) and antisymmetry (where
{d1 } ≺ {d2 } implies {d2 } ̸≺ {d1 }) are not generally assumed and depend on the domain. Symmetric
preference relations, such as {dcasualties } ≺ {dweapon } and {dweapon } ≺ {dcasualties }, can coexist for the
same focus case, indicating that a better value in one dimension can compensate for a worse value
in another. For instance, a high number of casualties (dcasualties ) may compensate for high weapon
sophistication (dweapon ) and vice versa. This symmetry may imply that dimensions are equivalent in
their influence on an outcome, acting as proxies for a more abstract notion. For example, dalert (security
alerted) and dmeasures (number of security measures) could be subcategories of a dimension dsecurity
(overall security preparedness). This implies a certain kind of equivalence. Thus our approach implicitly
allows for the drawing of abstract parallels similar to the factor hierarchies in CATO [1].
3.3. Arguing About Preference Relations
As mentioned, we do not assume the body of domain knowledge to be uncontested. While Schemes 1
and 2 provide an approach to assess whether the conditions of a compensation move have been met,
exceptions may be possible and premises can be contested. Exactly what kinds of attacks are possible
may depend on the domain, but in general we can state that attacks between arguments can be modelled
in a structured argumentation framework like ASPIC+ [13] or ABA [2].
For example, a domain expert may consider there to be another caveat for the preference relation
{dcasualties } ≺ {dweapon } besides dcasualties ( f ) ≥ 4, namely that it does not hold if the weapon sophistica-
tion is extremely high. The abstract placeholder Ψ could then simply refer to two separate thresholds for
this instance of PREF( f , Db , Dw , D):
PREF( f , Db , Dw , D):
ψ: dcasualties ( f ) > 4 ∧ dweapon ( f ) < ‘Extremely high’
===================
Conc: {dcasualties } ≺ {dweapon }
However, one might argue that it is more informative if exceptions are modelled explicitly as sepa-
rate arguments. The preference relation {dcasualties } ≺ {dweapon } would then be attacked by an excep-
tion argument detailing how it can be concluded from the fact that dweapon ( f ) ≮ ‘Extremely high’ that
{dcasualties } ≺ {dweapon } does not hold for f . This exception argument would have to be successfully
attacked in order for {dcasualties } ≺ {dweapon } to be usable in Scheme 1, perhaps by an exception to the
exception. For instance, the exception dweapon ( f ) ≮ ‘Extremely high’ might be considered irrelevant if
the number of casualties is sufficiently high, e.g. dcasualties ≥ 30. This second exception would attack
the first exception, thereby defending the preference relation from Scheme 2 and thus reinstating the
compensation move from Scheme 1. And so on for additional exceptions. This notion is illustrated in the
argument graph Figure 2.
We allow for chains of arguments about preference relations. Whether long, complex arguments are
always desirable is up to the domain experts themselves, based on what they deem appropriate for the
intended user. Our approach allows them to decide on how elaborately to justify the domain knowledge
used to justify ML predictions as they see fit. The goal is always to justify compensation moves in the
eyes of the user, who may or may not be a domain expert, in order to provide an appropriate level of
justification for ML predictions.
compensates(c, {dcasualties}, {dweapon}) compensates(c, {dcasualties}, {dweapon})
COMP(f,Db,Dw) COMP(f,Db,Dw)
{dcasualties} ≺ {dcasualties} ≺
Dw = {dcasualties} Db = {dweapon} Dw = {dcasualties} Db = {dweapon}
{dweapon} {dweapon}
EXCEPTION: EXCEPTION:
PREF(f,Db,Dw,D) PREF(f,Db,Dw,D)
dweapon(f) ≮ ‘Extremely high’ dweapon(f) ≮ ‘Extremely high’
dcasualties(f) ≥ 4 dcasualties(f) ≥ 4
EXCEPTION:
dcasualties(f) ≥ 30
Figure 2: An illustration of a how an exception to an exception can defend a preference relation and thus a
compensation move. Shaded boxes are not in the grounded extension, attacks are indicated by arrows.
Other types of argument could be valueable too, such as expert opinions (after Walton et al. [20])
on the basis of additional domain knowledge—based on professional experience, domain literature,
or statistical/data analysis (e.g. rule discovery). Of course, there are domain-specific reasons why a
preference relation is included in the first place, which suggests that conflicting opinions are possible.
Similarly to how we could allow chains of exceptions to argue about preference relations, experts may
decide that it is equally informative to explicitly model dissent between experts and perhaps to show how
the latest analysis, literature research or most senior expert wins the debate.
For example, the preference relation {dcasualties } ≺ {dweapon } could be attacked by an opinion stating
that it does not hold, based on the experience of the expert. This could itself be attacked by an opinion
on the basis of statistical analysis suggesting that even in scenarios where weapon sophistication was
extremely high, the number of casualties had a more significant impact on outcomes. Thus, the argument
from statistical analysis would then successfully defend the original preference relation.This example
only describes the approach generally and more detailed decisions regarding possible arguments and
attack types have to be made when it is implemented using a structured argumentation framework.
4. Conclusion
We have extended the XAI approach AF-CBA by adding a mechanism by which the justifications’
compensation moves are determined using arguments based on domain knowledge provided by domain
experts. This not only allows compensation moves to be informed by established domain rules, but also
communicates reasons for those compensation moves in terms that are likely to be familiar to domain
experts. We have implemented this mechanism as a secondary argument graph, which likewise can
be shown to the user as a justification. This secondary argument graph provides an avenue for future
extensions aimed at more in-depth justifications and possibly disputes about the domain knowledge itself.
Our extension of AF-CBA relies on argumentation schemes to capture defeasible reasoning patterns,
providing a foundation for persuasive justifications. Further extending these patterns within a structured
argumentation framework would enhance the sophistication of arguments, allowing arguments about
premisses or about the implied entailment of preference relations. This could, for instance, take the form
of arguments stemming from different sources of domain knowledge. The current paper is not on rule
discovery or information extraction, but those techniques could be integrated in a larger framework in
which any disagreements between sources of domain knowledge have to be resolved. Refining rules
(e.g. thresholds) is another aspect that deserves attention in future work. Another possible future work
direction is to take an experimental approach in the form of usability studies, which would allow us to
evaluate various design choices for AF-CBA from a user’s perspective.
References
[1] Vincent Aleven. Using background knowledge in case-based legal reasoning: A computational
model and an intelligent learning environment. Artificial Intelligence, 150(1-2):183–237, 2003.
[2] Andrei Bondarenko, P. M. Dung, R. A. Kowalski, and F. Toni. An abstract, argumentation-theoretic
approach to default reasoning. Artificial Intelligence, 93(1):63–101, 1997.
[3] Kristijonas Čyras, Ken Satoh, and Francesca Toni. Abstract Argumentation for Case-Based Rea-
soning. In Proceedings of the Fifteenth International Conference on Principles of Knowledge
Representation, 2016.
[4] Kristijonas Čyras, Ken Satoh, and Francesca Toni. Explanation for Case-Based Reasoning via
Abstract Argumentation. In Proceedings of COMMA 2016, pages 243–254. IOS Press, 2016.
[5] Kristijonas Čyras, David Birch, Yike Guo, Francesca Toni, Rajvinder Dulay, Sally Turvey, Daniel
Greenberg, and Tharindi Hapuarachchi. Explanations by arbitrated argumentative dispute. Expert
Systems with Applications, 127:141–156, 2019.
[6] Phan Minh Dung. On the acceptability of arguments and its fundamental role in nonmonotonic
reasoning, logic programming and n-person games. Artificial Intelligence, 2(77):321–357, 1995.
[7] Andrew R. Golding and Paul S. Rosenbloom. Improving accuracy by combining rule-based and
case-based reasoning. Artificial Intelligence, 87(1-2):215–254, 1996.
[8] Riccardo Guidotti, Anna Monreale, Salvatore Ruggieri, Franco Turini, Fosca Giannotti, and Dino
Pedreschi. A survey of methods for explaining black box models. ACM Computing Surveys, 51(5):
93:1–93:42, 2018.
[9] John Horty. Reasoning with dimensions and magnitudes. Artificial Intelligence and Law, 27(3):
309–345, 2019.
[10] John F. Horty. Rules and reasons in the theory of precedent. Legal Theory, 17(1):1–33, 2011.
[11] Zachary Lipton. The mythos of model interpretability. Communications of the ACM, 61:96–100,
2016.
[12] Tim Miller. Explanation in artificial intelligence: Insights from the social sciences. Artificial
Intelligence, 267:1–38, 2019.
[13] Sanjay Modgil and Henry Prakken. The ASPIC+ framework for structured argumentation: A
tutorial. Argument & Computation, 5(1):31–62, 2014.
[14] Joeri G.T. Peters and Floris J. Bex. Towards a Story Scheme Ontology of Terrorist MOs. In 2020
IEEE International Conference on Intelligence and Security Informatics (ISI), pages 1–6, 2020.
[15] Joeri G.T. Peters, Floris J. Bex, and Henry Prakken. Model- and data-agnostic justifications with
A Fortiori Case-Based Argumentation. In Proceedings of the 19th International Conference on
Artificial Intelligence and Law, pages 207–216, Braga, Portugal, 2023. Association for Computing
Machinery. ISBN 9798400701979.
[16] Henry Prakken. Dialectical proof theory for defeasible argumentation with defeasible priorities
(preliminary report). In John-Jules Ch. Meyer and Pierre-Yves Schobbens, editors, Formal Models
of Agents, Lecture Notes in Computer Science, pages 202–215, Berlin, Heidelberg, 1999. Springer.
ISBN 978-3-540-46581-2.
[17] Henry Prakken and Rosa Ratsma. A top-level model of case-based argumentation for explanation:
Formalisation and experiments. Argument & Computation, 13(2):159–194, 2022.
[18] Edwina L. Rissland and David B. Skalak. CABARET: Rule interpretation in a hybrid architecture.
International Journal of Man-Machine Studies, 34(6):839–887, 1991.
[19] Wijnand van Woerkom, Davide Grossi, Henry Prakken, and Bart Verheij. Hierarchical Precedential
Constraint. In Proceedings of the 19th International Conference on Artificial Intelligence and Law,
pages 333–342, Braga, Portugal, 2023. Association for Computing Machinery.
[20] Douglas Walton, Christopher Reed, and Fabrizio Macagno. Argumentation Schemes. Cambridge
University Press, 2008. ISBN 978-1-316-58313-5.