---
title: 'AA-CBR: Argumentation for Case-Based Reasoning'
url: https://www.emergentmind.com/topics/abstract-argumentation-for-case-based-reasoning-aa-cbr
type: topic
---

# AA-CBR: Argumentation for Case-Based Reasoning

Abstract Argumentation for Case-Based Reasoning (AA-CBR) is a case-based classification framework in which past cases, a default argument, and a new unlabelled case are compiled into an abstract argumentation framework, and the predicted outcome is obtained from the grounded extension by checking whether the default argument is accepted [2310.19607]. In this setting, cases are arguments, outcome disagreement and specificity induce attacks, and relevance determines which past cases the new case can neutralize as irrelevant [2007.05284]. AA-CBR originated as an argumentative reading of case-based reasoning, initially in legal reasoning and later in classification settings, and has since developed into a family of symbolic, preference-based, supported, and neuro-symbolic models [2107.06413].

## 1. Formal setting and core objects

AA-CBR is defined over a set \(X\) of characterisations and a binary outcome set \(Y\), written either as \(Y=\{+,-\}\) or \(Y=\{\delta,\bar{\delta}\}\). A casebase is a finite set \(D \subseteq X \times Y\), where each case \((x,y)\) pairs a characterisation with an outcome. A new case is represented as \((N_C,?)\), and a distinguished default argument \(\Delta=(\delta,\delta_o)\) or \((x^\bot,o_d)\) encodes the default outcome [2007.05284].

The framework assumes a partial order \(\succeq\) or \(\preceq\) over characterisations. In the standard reading, this order captures specificity or exceptionality: more specific cases are more exceptional, and more exceptional cases can defeat less exceptional cases with opposing outcomes. In the original subset-based instance, characterisations are sets of features and specificity is the superset relation, so \(x \succeq y\) iff \(x \supseteq y\) [2007.05284].

AA-CBR also assumes an irrelevance relation. In the regular version \(AA\text{-}CBR_{\succeq}\), irrelevance is tied directly to the same partial order: \(x_1 \not\sim x_2\) iff \(x_1 \not\succeq x_2\), and the default characterisation is the least element of the order [2007.05284]. This regularity condition is important because it makes specificity govern both case-to-case attacks and the query’s filtering of irrelevant cases.

A casebase is coherent when no two past cases share the same characterisation but have different outcomes. Coherence is technically significant in the standard theory, because many results on well-foundedness, concise subsets, and monotonicity variants are stated for coherent casebases [2310.19607].

## 2. Construction of the argumentation framework and prediction rule

AA-CBR builds a Dung-style abstract argumentation framework \(AF=(Args,Att)\). The arguments are all past cases in \(D\), the default argument, and the new case. Grounded semantics is used: the grounded extension \(G\) is the union of unattacked arguments and the arguments iteratively defended by them [2310.19607].

The attack relation has two parts. First, between past cases and the default argument, a case \((\alpha_C,\alpha_o)\) attacks \((\beta_C,\beta_o)\) when the outcomes disagree, \(\alpha_C \succeq \beta_C\), and there is no intermediate same-outcome case strictly between them in the partial order. This minimality condition ensures that attacks are induced only by the nearest relevant exceptions rather than by every more specific case of the opposite label [2007.05284].

Second, the new case attacks all past or default cases that are irrelevant to it. In the regular setting, \((N_C,?)\) attacks \((\beta_C,\beta_o)\) iff \(N_C \not\sim \beta_C\), which means that the new case explicitly removes cases that are not sufficiently related to its characterisation from the dialectical debate [2310.19607].

Prediction is determined by the status of the default argument in the grounded extension. If the default argument is accepted, the model predicts the default outcome; otherwise it predicts the non-default outcome. This gives AA-CBR a distinctive decision rule: it does not retrieve a single nearest precedent, and it does not aggregate by direct vote. Instead, it computes a grounded dialectical status over a graph of cases, exceptions, and irrelevance attacks [2007.05284].

A common defeasible-reasoning interpretation is that each case behaves like a defeasible rule of the form “if characterisation \(x\), then outcome \(y\),” while more exceptional cases defeat less exceptional incompatible ones. On this reading, the default argument is the most general defeasible rule, and the casebase organizes a hierarchy of exceptions around it [2510.00185].

## 3. Non-monotonicity, cautious monotonicity, and restricted casebases

AA-CBR is non-monotonic: adding cases can revoke earlier conclusions. More specifically, the regular model \(AA\text{-}CBR_{\succeq}\) is not cautiously monotonic, a property often regarded as desirable in non-monotonic reasoning [2007.05284]. In classifier terms, cautious monotonicity would mean that augmenting the dataset with conclusions already inferred by the classifier would not change other conclusions. The regular AA-CBR model fails this property.

This failure motivated a variant often denoted \(cAACBR\). Its central idea is to replace the full casebase with a restricted casebase consisting of cases that are “surprising” and “sufficient.” A case \((x,y)\) is surprising with respect to a dataset if removing it changes the prediction for \(x\); it is sufficient if adding it guarantees prediction \(y\) for \(x\); and it is includable if it is both surprising and sufficient [2107.06413].

The theory defines concise subsets as fixed points of includability and proves uniqueness for coherent datasets. An algorithm constructs the relevant restricted casebase by processing \(\succeq\)-minimal strata and adding only those cases whose outcome is not predicted by the current AF. The resulting model is equivalent to using \(AA\text{-}CBR_{\succeq}\) with the restricted casebase consisting of all “surprising” and “sufficient” cases in the original casebase [2107.06413].

This restriction has strong logical consequences. The cautiously monotonic variant is cautiously monotonic, cumulative, and rationally monotonic. It also supports a principled treatment of noise in incoherent casebases, because the constructed restricted casebase is coherent and excludes redundant or destabilizing cases. The monotonicity and noise-tolerance analysis was illustrated on a case study on the U.S. Trade Secrets domain [2107.06413].

This suggests a useful distinction within the literature. The original AA-CBR emphasizes the full argumentative structure induced by the casebase, whereas the cautious variant emphasizes a canonical core of cases needed to preserve stable inferential behavior.

## 4. Supports, spikes, and preference-sensitive extensions

A recurrent criticism of the original attack-only model is that not every case actively participates in classification. Supported Abstract Argumentation for Case-Based Reasoning, denoted sAA-CBR, addresses this by moving from ordinary abstract argumentation frameworks to bipolar argumentation frameworks with both attacks and supports [2507.04994]. In sAA-CBR, supports are added between same-label cases using the same exceptionality and minimality principles that govern attacks between opposite-label cases.

Supports are interpreted through supported and secondary attacks. The framework is translated back into an ordinary AF by adding original attacks, supported attacks, and secondary attacks, while excluding mediated attacks because those could create attacks from less exceptional to more exceptional cases and so violate a core AA-CBR design principle [2507.04994].

The principal structural target of sAA-CBR is the elimination of spikes. A spike is an argument for which there is no path to the default argument. In the attack-only setting, spikes are extraneous cases that never influence prediction. sAA-CBR proves that, when the default is the least exceptional case, there are no spikes. The paper presents this as overcoming the limitation of AA-CBR without trading off key model properties [2507.04994].

A different extension introduces preferences over how cases are compared. Preference-Based Abstract Argumentation for Case-Based Reasoning, AA-CBR-P, allows users to define multiple approaches to compare cases together with an ordering that specifies preference over those approaches. The model is proved to follow these preferences when making predictions, and the work argues that previous AA-CBR approaches are insufficient at expressing preferences over constituents of an argument. The paper reports an application to a medical dataset from a clinical trial evaluating differing assessment methods of patients with a primary brain tumour and states that the approach outperforms other interpretable machine learning models on that dataset [2408.00108].

Together, these extensions clarify that AA-CBR is not a single fixed architecture. The family now includes attack-only models, cautiously monotonic restricted-casebase models, bipolar models with supports, and preference-aware models with explicit control over competing comparison criteria.

## 5. Learning relevance and explanatory casebases

Although the original framework presupposes a partial order and irrelevance relation, later work showed that case relevance can be learned automatically from data. A technical report on legal prediction uses decision trees to learn a characterisation space for AA-CBR: decision-tree splits become binary features, cases are represented by the set of split conditions they satisfy, and the partial order is fixed as set inclusion in the form \(\preceq := \supseteq\) [2310.19607].

This construction preserves the standard AA-CBR prediction mechanism while replacing manual feature engineering with learned binary characterisations. The work studies three incoherence-resolution strategies—keep, removal, and majority—and reports that majority performs best. On COMPAS feature set C, the decision tree achieved \(67.48\pm1.56\) accuracy, while \(\mathsf{AA}\) and \(c\mathsf{AA}\) both achieved \(66.32\pm1.20\); on the Welfare dataset, the decision tree achieved \(99.6\pm0.1\), \(\mathsf{AA}\) achieved \(99.5\pm0.4\), and \(c\mathsf{AA}\) achieved \(99.6\pm0.2\) [2310.19607].

The same study emphasizes compactness. For COMPAS feature set C, the global models averaged \(143.0\pm184.9\) nodes for the decision tree, \(12.6\pm3.1\) for \(\mathsf{AA}\), and \(8.2\pm1.6\) for \(c\mathsf{AA}\). For Welfare, the corresponding averages were \(11.0\pm0.0\), \(7.8\pm4.3\), and \(4.6\pm0.5\) [2310.19607]. The local explanation device is the Arbitrated Dispute Tree, which renders a prediction as a proponent–opponent debate over accepted and rejected cases.

This line of work is important because it shows that AA-CBR need not be limited to hand-crafted legal factors. The learned-relevance approach retains the argumentative semantics, but learns a feature representation that induces the case ordering.

## 6. Neuro-symbolic and language-model developments

Recent work generalizes AA-CBR beyond purely symbolic case ordering. Gradual AA-CBR replaces crisp attack-only AFs with an edge-weighted Quantitative Bipolar Argumentation Framework \(\langle A,E,\tau,w\rangle\), where arguments have base scores, edges have real-valued weights, and gradual acceptability is computed through MLP semantics [2505.15742]. The model learns the outcome-determining debate structure simultaneously with neural-based feature extractors.

This gradual model differs from symbolic AA-CBR in several stated ways: it is capable of multi-class classification, automatic learning of feature and data point importance, assigning uncertainty values to outcomes, using all available data points, and not requiring binary features. Empirically, it performs comparably to neural networks while significantly outperforming existing AA-CBR formulations: on Mushroom, accuracy is \(0.48\) for AA-CBR, \(0.84\) for ANNA, \(0.99\) for the neural network, and \(0.98\) for Gradual AA-CBR; on Glioma, the corresponding values are \(0.67\), \(0.67\), \(0.85\), and \(0.86\); on Iris, the neural network and Gradual AA-CBR both achieve \(0.97\) accuracy [2505.15742].

Another neuro-symbolic line integrates object-centric perception with symbolic reasoning. Slot Attention Argumentation for Case-Based Reasoning, SAA-CBR, uses a Slot Attention component to extract object-level attributes from images and then classifies with AA-CBR over symbolic characterisations built from those attributes. The framework introduces super-features, count-based characterisations, casebase reduction via k-means and thresholding, a One-Vs-Rest strategy for multi-class classification, and an application of Supported AA-CBR [2510.00185]. On CLEVR-Hans3 test data, SAA-CBR achieved \(75.13\%\) accuracy and \(74.77\%\) macro-F1; on modified CLEVR-Hans7 test data, it achieved \(62.87\%\) accuracy and \(63.19\%\) macro-F1, ranking second to NS-CL on both datasets [2510.00185].

A further development addresses non-factorized case bases with language models. Argumentative agentic models for case-based reasoning, AAM-CBR, retain the standard AA-CBR symbolic layer but use language-model agents to determine case coverage and extract factors from textual previous cases relative to a new case. This allows factor-based reasoning without exposing or preprocessing previous cases. The reported result is that AAM-CBR surpasses a single-prompt baseline only when the new case contains a richer set of factors, indicating that symbolic reasoning remains important as factor complexity increases [2512.12656].

These developments show that contemporary AA-CBR research no longer treats symbolic argumentation and learned representation as mutually exclusive. Instead, AA-CBR increasingly functions as the reasoning layer in larger neuro-symbolic pipelines.

## 7. Applications, interpretability, and persistent limitations

AA-CBR has been discussed as applicable in legal reasoning, image classification, sentiment analysis of text, and prediction of the passage of bills in the UK Parliament, and later work has instantiated it in legal datasets, synthetic visual reasoning benchmarks, medical data, and synthetic credit decision data [2007.05284]. The common thread is that AA-CBR yields case-based explanations whose basic units are concrete examples rather than latent parameters.

The explanatory value of the framework follows directly from its structure. In AA-CBR, each accepted or rejected argument is a case, attacks show which cases overrule which others, and the default argument makes the prediction criterion explicit. In learned-relevance legal settings, Arbitrated Dispute Trees provide local debates over the status of the default argument; in object-centric image classification, explanations are grounded in symbolic object attributes such as shape, colour, size, material, and position rather than in opaque activations [2310.19607].

At the same time, several limitations recur across the literature. The original AA-CBR is binary; multi-class prediction requires strategies such as One-Vs-Rest or a gradual target-argument construction [2510.00185]. The classical framework depends on a suitable partial order and irrelevance relation, and their design is domain-sensitive [2107.06413]. Attack-only AA-CBR can contain spikes, which means some cases never affect any prediction [2507.04994]. In SAA-CBR, feature combination can become combinatorial, casebase reduction is generic rather than argumentation-aware, and AF construction is \(O(n^3)\) in the number of cases [2510.00185]. In Gradual AA-CBR, performance is sensitive to initial weights, and successful learning rates varied substantially across datasets [2505.15742].

A common misconception is that AA-CBR is simply a nearest-neighbour method with argumentative language layered on top. The formal definitions suggest a different picture: the core computation is grounded acceptance in an argumentation framework mined from a casebase, with exceptionality, irrelevance, default reasoning, and sometimes supports or preferences all affecting the result. A second misconception is that every stored case necessarily matters in the original model; the literature on spikes shows that this is not generally true, and later variants were introduced precisely to change that [2507.04994].

AA-CBR therefore occupies a specific position in the broader landscape of interpretable machine learning and formal argumentation. It is a classifier family in which prediction is inseparable from dialectical structure, and current research is extending that structure rather than abandoning it: by restricting casebases for monotonicity, by adding supports and preferences for expressiveness, and by coupling symbolic debate with learned representations for scale and broader data modalities [2408.00108].

Source: https://www.emergentmind.com/topics/abstract-argumentation-for-case-based-reasoning-aa-cbr