Source-linked AI summary
Contrastive Explanations in Quantitative Bipolar Argumentation Frameworks
Xiang Yin, Nico Potyka, Antonio Rago, Francesca Toni
TL;DR
Existing QBAF explanations focus on single topic arguments, leaving the contrast between two topic arguments unexplained. The paper introduces contrastive attribution functions and three instantiations, then demonstrates their use in healthcare and bias identification, including revealing a race-related contrast in an MLP prediction.
Problem
Existing QBAF explanation methods mainly explain one topic argument and overlook why it was selected rather than another argument.
Method
The paper defines CAFs for strength differences, establishes desirable properties, and studies removal-, gradient-, and Shapley-based instantiations.
Results
Contrastive explanations reveal race-related bias: race has attribution 0.2764 for Medium rather than Low, the largest among eight inputs, versus −0.0175 for Medium alone.
Takeaways & Limitations
Contrastive CAFs make discriminative influences visible in QBAF reasoning and can expose bias that is not prominent in individual argument explanations.
Takeaways & Limitations
The paper identifies extending the approach to rankings, conducting user studies, and exploring further application domains as future work.
Abstract
from arXiv · showhide
Argumentation frameworks are useful tools for representing and reasoning with information in a variety of settings, e.g. in supplementing AI models as they perform classification tasks, with a notable benefit of providing additional explainability. In this paper, we introduce contrastive explanations for Quantitative Bipolar Argumentation Frameworks (QBAFs), one such formalism. Unlike most existing explanations for QBAFs, which explain the reasoning outcome of a single argument of interest (i.e. a topic argument), contrastive explanations explain the difference between two topic arguments. We introduce a general form of contrastive attribution functions (CAFs) and establish a set of general properties they should satisfy. We introduce CAFs based on removal, gradients and Shapley-values, and study their properties. Finally, to illustrate contrastive explanations, we demonstrate their usefulness in healthcare and bias identification settings.
1 Introduction
The paper motivates contrastive explanations for QBAFs because existing methods explain one topic argument, while users often need to know why one prediction holds rather than another. It introduces attribution-based CAFs and situates them within QBAF semantics and existing attribution methods.
- Motivation: QBAFs represent conflicting and supporting information for reasoning tasks where explainability may otherwise be lacking.They are used in settings including image classification, recommendation, and claim verification.
- Motivation: Contrastive explanations identify reasons distinguishing a prediction P from a contrast case Q rather than listing every explanatory reason.For zebra, stripes distinguish zebra from horse, whereas being a herbivore distinguishes zebra from tiger.
- Research gap: Existing QBAF explanations mainly explain the strength of one topic argument and therefore overlook why the alternative argument was not selected.This motivates explaining the difference between the strengths of two topic arguments.
- Approach: The proposed attribution-based CAFs assign scores to non-topic arguments according to their influence on the strength gap between two topic arguments.Comparing scores identifies discriminative arguments without manually inspecting all reasoning paths.
- Contributions: The paper introduces general CAF properties, studies removal-, gradient-, and Shapley-based methods, and illustrates applications in healthcare and bias detection.The stated contributions span formal properties, concrete methods, and application demonstrations.
3 Contrastive Explanations
Contrastive attribution functions quantify how arguments affect the relative strength of two topic arguments. The paper proposes properties governing their signs, symmetry, calibration, and neutrality.
- Contrastive attribution functions: A CAF assigns each non-topic argument a real-valued influence on the relative strength of two topic arguments.For α ⪰ β, Φα⪰β maps arguments other than α and β to real numbers.
- Interpretation: Positive influence means an argument contributes more favourably to α than to β, while negative influence favours β relative to α.Influence is neutral when the argument contributes equally to both topic arguments.
- Desirable properties: Antisymmetry requires reversing the topic order to negate each argument’s contrastive influence.It also implies zero influence when both topic arguments are identical.
- Desirable properties: Calibration preserves a reference attribution when an argument influences one topic but not the other, while inverse calibration handles the reversed case.If the reference values are a and 0, the CAF value is a; if they are 0 and b, it is −b.
- Desirable properties: Neutrality requires zero contrastive influence when an argument has equal attribution to both topic arguments.Monotonicity additionally requires positive contrastive influence when a reference attribution favors α over β, and negative influence in the reverse case.
5 Derived CAFs
Derived CAFs construct contrastive influence by combining individual attributions for two topic arguments, with subtraction providing a principled and characterized combination.
- Derived CAFs: Derived CAFs combine individual attribution values for two topic arguments through a binary function.For each non-topic argument γ, the CAF applies f to its attributions for α and β.
- Properties: Subtraction-derived CAFs satisfy antisymmetry, calibration, neutrality, and monotonicity.
- Properties: Subtraction-derived CAFs also satisfy additivity, allowing contrastive effects to decompose through an intermediate topic argument.
- Characterization: Under modular semantics, antisymmetry, calibration, and additivity characterize subtraction-derived CAFs for plausible attribution functions.The characterization applies when the reference attribution function is plausible.
- Concrete CAFs: The paper instantiates derived CAFs from removal-based, gradient-based, and Shapley-based attribution functions.These methods measure removal effects, local base-score sensitivity, and average marginal contributions, respectively.
6 Computing Subtraction-Derived CAFs
Additivity enables efficient computation of contrastive preferences across many topic arguments, while additional properties distinguish CAF interpretations such as counterfactuality and local sensitivity.
- Efficient computation: Naively evaluating all ordered preferences among T topic arguments requires T · (T −1) = O(T^2) CAF calls.
- Efficient computation: Additivity reduces the required CAF calls to O(T) through a dynamic programming algorithm.The algorithm reconstructs nonadjacent contrasts from adjacent ones.
- Efficient computation: If each CAF call takes O(C) time, the dynamic programming algorithm runs in O(T · C + T^2) time.
- Method-specific properties: Counterfactuality links an argument’s attribution sign to how removing it changes the topic-strength difference.
- Method-specific properties: Local faithfulness links attribution signs to the local effect of increasing an argument’s base score on the strength gap.
- Method-specific properties: The three additional properties guide CAF choice: removal supports counterfactual interpretations, gradients support robustness analysis, and Shapley values support marginal-contribution explanations.
8 CAFs for Healthcare
The healthcare example uses a layered QBAF to compare treatment arguments and shows how Shapley-based CAFs identify symptoms and diagnoses that distinguish antiviral from antibiotic therapy.
- Healthcare QBAF: The healthcare QBAF organizes symptoms, diagnoses, and treatments into three hierarchical layers connected by attack and support relations.
- Healthcare QBAF: All arguments receive base score 0.5, while QE semantics yields diagnosis strengths of 0.75, 0.6, and 0.4 for viral pneumonia, bacterial pneumonia, and asthma.
- Contrastive explanations: Shapley-based contrastive explanations identify Positive PCR Test as the largest positive influence on antiviral therapy over antibiotic therapy.Its paths support antiviral therapy and weaken antibiotic therapy through the diagnosis layer.
- Contrastive explanations: Viral pneumonia has the strongest positive influence, while bacterial pneumonia has the strongest negative influence on the treatment contrast.Their direct support and attack relations favor opposite treatments.
- Individual versus contrastive explanations: Asthma is nearly neutral contrastively because it negatively influences both treatments, whereas elevated WBC count becomes strongly negative through its positive influence on antibiotic therapy.
- Property illustration: 0.1196 is the final antiviral–antibiotic strength difference, matching the adjusted sum of non-topic and cross-topic Shapley contributions.The non-topic attributions sum to 0.1123 and the cross-topic component is 0.0074.
9 CAFs for Bias Detection
The paper applies gradient-based contrastive explanations to controlled race bias in a classifier, showing that contrastive attributions expose race-related influence that individual explanations can obscure. Across 141 instances, race becomes substantially more prominent in contrastive explanations.
- Method: Gradient-based contrastive attributions are used to uncover bias in classification tasks while avoiding Shapley-based combinatorial calculations.The analysis uses local sensitivity in an MLP represented as a QBAF.
- Experimental setup: 70% of African-American Low labels were randomly relabeled as Medium, and around 70% of Medium labels as High, while Caucasian labels were unchanged.This controlled modification was designed to make the classifier learn a race-related bias.
- Experimental setup: Changing race from African-American to Caucasian changed one correctly classified defendant’s prediction from Medium to Low.All other input features were held fixed, and the MLP was represented as a QBAF.
- Single-instance result: Race’s individual attribution was −0.0175, but its contrastive attribution for Medium rather than Low was 0.2764, the largest among inputs.The individual attribution was smallest in absolute magnitude, whereas the contrastive attribution revealed substantial race influence.
A Proofs
This section establishes structural properties of contrastive attribution functions and derives consequences for subtraction-based CAFs. It also proves an efficient algorithm for computing all pairwise topic contrasts and documents counterexamples for some methods.
- Properties: Antisymmetry requires reversing the two topic arguments to negate the contrastive attribution.It also implies that a contrast of an argument with itself is zero.
- Properties: Subtraction-derived CAFs satisfy antisymmetry, calibration, neutrality, monotonicity, and additivity.These properties follow directly from subtracting the two individual attribution values.
- Characterization: Under modular gradual semantics, a plausible CAF satisfying antisymmetry, calibration, and additivity must equal the subtraction-derived CAF.The result uses independence under the addition of isolated arguments.
- Algorithm: Additivity lets Algorithm 1 compute all pairwise CAF values using O(T) CAF calls and overall time O(T · C + T^2).Here C is the time required for one CAF computation.
- Counterexamples: Gradient- and Shapley-based CAFs may violate Counterfactuality: a positive attribution can occur even when removing an argument increases the strength difference.In the example, the original difference is 11/32, while the difference after removal is 3/8.
- Counterexamples: Removal- and Shapley-based CAFs may violate Local Faithfulness when a positive attribution accompanies a decreasing strength difference after a small base-score increase.The example reports a removal-based attribution of 3/64 and a Shapley-based attribution of 7/192.
B Additional Contrastive Explanations in Section 8
Figure 5 presents gradient-based contrastive and individual explanations for treatment selection, using colors to distinguish positive and negative influence.
- Figure 5 uses green bars for positive influence and red bars for negative influence in treatment-selection explanations.
C Computing Environment
The experiments were run locally on a Windows 11 laptop using CPU-only computation and a specified Python scientific-computing stack.
- Hardware and environment: All computations were performed locally on a Microsoft Windows 11 Home laptop.
- Hardware and environment: The laptop used an Intel Core Ultra 5 225H CPU with 14 cores and 32 GB of RAM.
- Software: The software environment included Python 3.11.9, PyTorch 2.13.0, NumPy 2.4.4, pandas 3.0.3, scikit-learn 1.9.0, and Matplotlib 3.10.8.