Contrastive Explanations in Quantitative Bipolar Argumentation Frameworks
Xiang Yin, Nico Potyka, Antonio Rago, Francesca Toni
cs.AI
Quick verdict
Adds contrastive explanations to QBAFs via subtraction of single-topic attributions; clean axiomatization, thin empirical validation (planted-bias COMPAS).
Summary
The paper introduces contrastive attribution functions (CAFs) for Quantitative Bipolar Argumentation Frameworks (QBAFs): instead of explaining the strength of a single topic argument, a CAF scores every non-topic argument by its influence on the strength gap between two topic arguments, answering 'why α rather than β?'. Building on existing attribution functions (removal-, gradient-, and Shapley-based), the authors derive CAFs by subtraction and prove they satisfy antisymmetry, ϕ-calibration, ϕ-neutrality, and ϕ-monotonicity. A characterization result (Proposition 7) shows that, over plausible attribution functions under modular semantics, antisymmetry plus ϕ-calibration plus additivity force a CAF to be exactly the subtraction-derived one. Additivity also yields a dynamic-programming algorithm needing O(T) CAF calls for T topics. Three method-specific properties separate the variants: counterfactuality holds only for removal-based CAFs, local faithfulness only for gradient-based CAFs, and cross-topic-adjusted efficiency only for Shapley-based CAFs, each demonstrated with explicit DF-QuAD counterexamples. Applications are illustrative: a hand-built healthcare QBAF where Shapley CAFs identify the PCR test as the most influential symptom discriminating antiviral from antibiotic therapy, and a bias-detection experiment on COMPAS where an MLP is deliberately biased by relabeling 70% of African-American instances; contrastive attribution raises race's mean importance rank from 5.16 to 2.62 over 141 instances. Limitations: no user study, no external XAI baselines, descriptive statistics only, and code is not released.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Contrastive attribution functions for QBAFs are obtained from any reference attribution function by pairwise subtraction and inherit antisymmetry, ϕ-calibration, ϕ-neutrality and ϕ-monotonicity (Props 4, 8).
- Under modular gradual semantics, antisymmetry plus ϕ-calibration plus additivity uniquely characterize subtraction-derived CAFs over plausible attribution functions (Prop 7).
- The removal-, gradient- and Shapley-based CAFs are axiomatically separated: only removal satisfies Counterfactuality, only gradient satisfies Local Faithfulness, and only Shapley satisfies Cross-topic-adjusted Efficiency (Props 10-12).
- Additivity enables computing all pairwise CAFs for T topic arguments with only O(T) CAF calls via dynamic programming (Algorithm 1, Prop 9).
- On a deliberately biased COMPAS classifier, contrastive attribution makes the race feature far more prominent (mean rank 5.16→2.62 over 141 instances) than individual attribution, exposing the planted bias.
Strengths
Clean uniqueness characterization
Prop 7 proves that under modular gradual semantics, over plausible attribution functions, antisymmetry + ϕ-calibration + additivity force a CAF to be exactly the subtraction-derived one; full proofs are supplied, including the non-trivial Shapley plausibility weight-merge argument in supplementary Proposition 6.
Axiomatic separation with hand-checkable counterexamples
Props 10-12 use fully specified 4-argument QBAFs under DF-QuAD with exact rationals (e.g., σQ(α)=17/32, σQ(β)=3/16; both ΦG and ΦS equal 1/8 > 0 while removal raises the gap from 11/32 to 3/8), so every claimed property failure is verifiable by hand.
Practical computation reduction
Demonstrated diagnostic value of the contrast
Honest scoping
Weaknesses
Self-referential empirical validation
No external XAI baselines
Purely descriptive statistics on a conditioned sample
Formal-statement/proof mismatch in Definition 9
Cognitive-adequacy claim untested
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This method paper (arXiv 2609.02399, cs.AI) by Yin, Potyka, Rago and Toni introduces contrastive attribution functions (CAFs) for Quantitative Bipolar Argumentation Frameworks (QBAFs). Where prior QBAF explanation work — AAEs (Yin et al. 2023), CE-QArg (Yin et al. 2024a), contribution functions (Kampik et al. 2024) — explains the strength of a single topic argument, CAFs quantify each non-topic argument's influence on the strength gap between two topic arguments, formalizing the "why P rather than Q" question from the contrastive-explanation literature (Lipton 1990; Miller 2019/2021). Central claims: (1) CAFs derived from any reference attribution function by subtraction (Def 8) satisfy antisymmetry, ϕ-calibration, ϕ-neutrality and ϕ-monotonicity (Prop 4); (2) under modular semantics, antisymmetry + ϕ-calibration + additivity uniquely characterize subtraction-derived CAFs over plausible AFs (Prop 7); (3) the removal-, gradient- and Shapley-based CAFs are separated by Counterfactuality, Local Faithfulness and Cross-topic-adjusted Efficiency respectively (Props 10-12); (4) additivity yields an O(T)-call dynamic program (Alg 1, Prop 9); (5) contrastive attribution surfaces planted race bias that individual attribution obscures (Sec 9).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is a contrastive explanation in a Quantitative Bipolar Argumentation Framework?
It is a score assigned to each non-topic argument measuring how much it influences the strength difference between two topic arguments — e.g. why the model prefers zebra over horse. The paper computes it by subtracting the argument's two individual attribution scores, so an argument that influences both topics equally receives zero.
What are the three types of contrastive attribution functions proposed?
Removal-based (how the strength gap changes when an argument is deleted; satisfies counterfactuality), gradient-based (sensitivity of the gap to small base-score changes; satisfies local faithfulness), and Shapley-based (average marginal contribution across coalitions; satisfies cross-topic-adjusted efficiency). Each is separated from the others by an axiom with an explicit counterexample under DF-QuAD semantics.
What did the COMPAS bias experiment show?
On an MLP deliberately biased by relabeling 70% of African-American instances, the race feature ranked last (mean rank 5.16 of 8) under individual attribution but second (2.62) under contrastive attribution across 141 test instances — showing the contrastive view surfaces a planted bias that individual attributions hide.
Is code or data available for this paper?
No code release is mentioned. The COMPAS dataset is public (ProPublica/Kaggle), and the healthcare QBAF plus all formal counterexamples are fully specified in the text, so reimplementation is possible; however, training details such as optimizer, epochs, and random seeds are not given.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 appear only as captions in the extracted text; the bar-chart values (Figs 3-4) are known solely from the prose numbers quoted around them, so figure-level readings rely on the authors' text.
- The review is based solely on the provided text; references marked 'to appear' (e.g., Yin et al. 2026a) could not be externally verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.