sAA-CBR: Bipolar Refinement in Case-Based Reasoning

• sAA-CBR is a binary classification model that extends AA-CBR by incorporating same-label supports to remove inactive 'spikes'.
• The model builds a bipolar argumentation framework using both attacks and supports, ensuring evidence propagates through chains of agreement.
• Grounded semantics and a default argument rule maintain minimal, directional decision-making while preserving interpretability.

Supported Abstract Argumentation for Case-Based Reasoning (sAA-CBR) is a binary classification model in which past cases engage in debates by arguing in favour of their labelling and attacking or supporting those with opposing or agreeing labels. It extends the earlier AA-CBR framework by adding supports between same-label cases, with the explicit purpose of removing spikes: cases that are present in the casebase but do not participate in any reasoning path leading to the prediction. The model is presented as preserving the interpretability and case-based argumentative structure of AA-CBR while ensuring that previously extraneous cases can contribute to the debate (Gould et al., 7 Jul 2025).

1. Position within argumentation-based case-based reasoning

sAA-CBR was introduced as a direct extension of AA-CBR. In AA-CBR, each labelled past example is treated as an argument, cases with opposing labels attack one another, the default case argues for a default outcome, and grounded argumentation semantics determine whether the default argument survives. This yields an interpretable, case-based decision process, but the attack-only structure can produce disconnected cases under the minimality condition used to define attacks. Those disconnected cases are the spikes that motivate sAA-CBR (Gould et al., 7 Jul 2025).

The central change is conceptually narrow but structurally important: sAA-CBR adds supports between cases with the same label. A more exceptional case that agrees with a less exceptional one can therefore reinforce it, allowing evidence to propagate through chains of agreement rather than remaining inert. This suggests that sAA-CBR should be understood not as a replacement of AA-CBR’s argumentative logic, but as a bipolar refinement of it.

Aspect	AA-CBR	sAA-CBR
Relations between cases	Attacks only	Attacks and supports
Structural issue	Can contain spikes	Contains no spikes
Same-label interaction	No reinforcement	More exceptional same-label cases support less exceptional ones

The model remains within the case-based reasoning tradition because prediction is still derived from relations among labelled precedents and a new case. Its novelty lies in making agreement, not only disagreement, part of the formal debate.

2. Formal setting and representation

sAA-CBR is defined for binary classification with label set

$Y=\{\delta,\bar{\delta}\},$

where $\delta$ is the default outcome and $\bar{\delta}$ is its opposite (Gould et al., 7 Jul 2025).

A casebase is a finite set of labelled cases

$D \subseteq X \times Y,$

where each case has a characterisation $x \in X$ and a label $y \in Y$. Classification concerns a new unlabeled case

$N=(x_N,y_{?}),$

whose label must be inferred (Gould et al., 7 Jul 2025).

The framework assumes two additional structures over $X$. The first is a partial order $\succcurlyeq$, used to formalize exceptionality. The second is a binary irrelevance relation $\nsim$, used to determine whether a past case should be considered relevant to the new case. The intended reading is that more exceptional cases can attack or support less exceptional ones, while the new case can exclude irrelevant past cases from the debate (Gould et al., 7 Jul 2025).

The argumentative structure is built from three components: all labelled past cases, a default argument $\delta$0, and the new case $\delta$1. The resulting bipolar argumentation framework therefore contains a set of arguments together with an attack relation and a support relation. In this representation, cases are not merely stored precedents; they are active argumentative objects whose mutual relations determine the classification outcome.

3. Construction of attacks and supports

Attacks in sAA-CBR retain the core logic of AA-CBR. For two past cases $\delta$2 and $\delta$3, an attack requires label disagreement and a relation of exceptionality subject to a minimality condition. Concretely, $\delta$4 attacks $\delta$5 when $\delta$6 and either $\delta$7 is more exceptional than $\delta$8 with no intermediate case between them in the exceptionalism ordering, or the two cases are equivalent in that ordering. The minimal exceptional attack condition is stated as

$\delta$9

and

$\bar{\delta}$0

The symmetric case is

$\bar{\delta}$1

The new case attacks a past case $\bar{\delta}$2 whenever that case is irrelevant to it: $\bar{\delta}$3 All of these clauses are inherited from the argumentative case-based formulation that sAA-CBR extends (Gould et al., 7 Jul 2025).

The novel element is the support relation. A case $\bar{\delta}$4 supports case $\bar{\delta}$5 if they have the same label and $\bar{\delta}$6 is more exceptional than $\bar{\delta}$7 under the same minimality condition: $\bar{\delta}$8

$\bar{\delta}$9

Supports therefore mirror attacks structurally: attacks encode minimal exceptional disagreement, whereas supports encode minimal exceptional agreement (Gould et al., 7 Jul 2025).

This symmetry is the main conceptual shift of sAA-CBR. In AA-CBR, a more exceptional same-label case could be present yet unable to influence the outcome. In sAA-CBR, that case can reinforce a less exceptional ally, making chains of agreement part of the reasoning process.

4. Semantics and prediction

To compute acceptance, sAA-CBR translates the bipolar framework into an ordinary abstract argumentation framework using supported attacks and secondary attacks. A supported attack from $D \subseteq X \times Y,$0 to $D \subseteq X \times Y,$1 exists when there is a sequence

$D \subseteq X \times Y,$2

with

$D \subseteq X \times Y,$3

such that the path consists of supports until some argument that attacks $D \subseteq X \times Y,$4. A secondary attack is defined dually: there is a sequence

$D \subseteq X \times Y,$5

with

$D \subseteq X \times Y,$6

such that the path consists of attacks until some argument that supports $D \subseteq X \times Y,$7 (Gould et al., 7 Jul 2025).

The paper emphasizes that these derived attacks preserve the directional principle of the model: they still go from more exceptional to less exceptional cases. This matters because other bipolar notions, such as mediated attacks, could create attacks from less exceptional to more exceptional cases, which would violate the logic of case-based reasoning as formulated here (Gould et al., 7 Jul 2025).

Once the derived attack graph is obtained, sAA-CBR applies grounded semantics. The grounded extension is

$D \subseteq X \times Y,$8

where $D \subseteq X \times Y,$9 is the set of unattacked arguments and $x \in X$0 is the set of arguments defended by $x \in X$1. A set $x \in X$2 defends an argument $x \in X$3 if every attacker of $x \in X$4 is itself attacked by some member of $x \in X$5 (Gould et al., 7 Jul 2025).

The prediction rule is then: $x \in X$6 and otherwise

$x \in X$7

Classification is therefore still determined by whether the default argument is accepted. The addition of supports changes how the debate unfolds, but not the final decision mechanism (Gould et al., 7 Jul 2025).

5. Spikes, the illustrative example, and the no-spikes result

A spike is defined as an argument with no path to the default argument. In AA-CBR, spikes arise because the minimality condition on attacks can isolate cases from the reasoning chain that ultimately determines the label. Such cases remain in the casebase but are extraneous to the prediction (Gould et al., 7 Jul 2025).

The paper’s running example is a diet-classification scenario. The default outcome is unhealthy diet $x \in X$8, represented by the default argument $x \in X$9. The new patient $y \in Y$0 has features $y \in Y$1, and the past cases are

$y \in Y$2

In AA-CBR, $y \in Y$3 attacks $y \in Y$4, $y \in Y$5 attacks $y \in Y$6, and $y \in Y$7 is a spike because it does not lie on a path to the default outcome and contributes nothing to the classification. In the paper’s AA-CBR illustration, the accepted arguments lead to predicting a healthy diet $y \in Y$8. In sAA-CBR, by contrast, $y \in Y$9 supports $N=(x_N,y_{?}),$0, that support induces a supported attack from $N=(x_N,y_{?}),$1 to $N=(x_N,y_{?}),$2, and the prediction flips to unhealthy $N=(x_N,y_{?}),$3 (Gould et al., 7 Jul 2025).

The principal formal result is the no-spikes theorem. If $N=(x_N,y_{?}),$4 is the least element of the exceptionalism order $N=(x_N,y_{?}),$5, then the bipolar argumentation framework generated by sAA-CBR contains no spikes. The proof idea is chain-based: because the default is least exceptional, every argument can be connected through minimal attacks or supports toward it. A plausible implication is that sAA-CBR eliminates a form of structural redundancy that was intrinsic to the attack-only version of the model (Gould et al., 7 Jul 2025).

6. Preserved properties, scope, and later use

The paper explicitly states that sAA-CBR removes spikes without sacrificing key AA-CBR principles. These preserved properties are: minimality, since relations are still created only between minimally different cases; directionality by exceptionalism, since attacks and supports go from more exceptional to less exceptional cases; interpretability, because the model remains a visible case-based debate; and the binary decision mechanism, because prediction still depends on whether the default argument belongs to the grounded extension (Gould et al., 7 Jul 2025).

The scope of the original sAA-CBR paper is theoretical rather than empirical. It does not report empirical benchmark results. Instead, it provides the formal definition of the model, an illustrative example, the theoretical proof of the no-spikes property, and a discussion of future work, including empirical evaluation of classification performance and exploration of supports in preference-based or cumulative AA-CBR (Gould et al., 7 Jul 2025).

A common source of confusion is the similarity of names between sAA-CBR and SAA-CBR. The latter refers to Slot Attention Argumentation for Case-Based Reasoning, a later neuro-symbolic image-classification pipeline. That later system integrates Slot Attention with AA-CBR, introduces a One-Vs-Rest strategy for multi-class classification, and experiments with Supported AA-CBR as one component, treating whether to enable supports as a hyperparameter. SAA-CBR is therefore an application context in which Supported AA-CBR is used, not a synonym for sAA-CBR itself (Gaul et al., 30 Sep 2025).

Within the current literature, sAA-CBR is best understood as a formal refinement of argumentation-based case-based classification: it preserves the original default-based grounded decision rule, enriches the structure of inter-case relations with same-label support, and establishes the absence of spikes as a theorem rather than as an empirical tendency.