| paper | Axiomatic Aggregations of Abductive Explanations |
| authors | — |
| venue | AAAI 2024 |
| filed under | unclassified |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1) |
| judge confidence | high |
| authors would recognise it | yes |
Theorems 1–3 characterize aggregators axiomatically, while Proposition 4 is an axiomatic impossibility result; none asserts a complexity or algorithmic classification. The proposed mass-weighted Deegan–Packel expectation is a plausible high-multiplicity extension that the authors might recognize, but it is not a mirror of a named computational theorem. The mandatory computational-anchor test therefore fails.
fails bit a — no named computational result to mirror
The population only reweights unchanged local AXp explanations, and the population formula requires a new mass-linearity assumption absent from Theorem 2.
fatal: True
The proposed lift covers the Deegan–Packel aggregator and could analogously lift the Holler–Packel and Responsibility indices, but it does not cover a named computational result; Proposition 4 and the empirical results remain outside the mirror.
The strongest honest case is qualified: under the programme’s strict rule, this paper has no eligible computational anchor. Theorem 1, Theorem 2, and Theorem 3 are proved here, but they are uniqueness characterizations of aggregators, not results placing a problem in P, NP-hardness, FPT, or any other complexity class. Proposition 4 is also proved here, but is an axiomatic impossibility result. The statement that abductive explanations are generally intractable is only an unnumbered claim attributed to Marques-Silva and Ignatiev (2022). Thus I cannot honestly present this paper as having established a computational result whose continuous analogue is being mirrored.
The best positive near-miss is a population lift of Theorem 2, the Deegan–Packel characterization. I would call the resulting problem Population Deegan–Packel Aggregation\(_\infty\).
Fix a finite feature set \(N\), finite feature domains \(X_i\), and one binary classifier \(f:\prod_iX_i\to\{0,1\}\), used by the whole population. A type is a complete feature vector \(t\in X\): agents with the same \(t\) are indistinguishable to the classifier and to the explanation procedure. The instance gives a finite support \(T=\{t_1,\ldots,t_\tau\}\) and rational masses \(\mu_t\ge0\) summing to one. For each type, let \(M(t,f)\) be its set of AXp’s and \(M_i(t,f)\) the AXp’s containing feature \(i\).
The required output is the exact population-level importance vector \(z\in\mathbb{Q}^{|N|}\), where
\[ z_i = \sum_{t\in T}\mu_t \sum_{S\in M_i(t,f)}\frac{1}{|S|}. \]
Equivalently, \(z_i\) is the expected Deegan–Packel importance of feature \(i\) for a uniformly sampled member of the society. The decision variable is the reported feature-importance vector, or its induced top-\(k\) feature set; the objective is to compute the exact mass-weighted explanation of the population. A fuller version can return the per-type vectors as well as their aggregate.
This is a genuine population continuization. Think of millions of loan applicants assessed by one fixed policy, with age, purpose, credit score, income, and balance discretized into a few thousand recurring profiles. A society with \(n_t\) applicants of profile \(t\) becomes \(\mu_t=n_t/\sum_u n_u\). Here \(N\) may be millions while \(\tau\) is in the thousands: the relevant quantity is the fraction of applicants in each profile, not the identity of each applicant. The same construction applies to insurance claims, benefits screening, medical triage, or any large administrative system with repeated structured records.
The mirror is plausible because it preserves the paper’s central object: AXp’s are still defined for the same classifier and the same data-point types, and the Deegan–Packel score is unchanged locally. We only ask for the natural population summary of the scores that the paper already proposes. Theorem 2, proved in this paper, supplies the local justification: with the corresponding mass-additivity and population-symmetry axioms, the natural lifted aggregator is the mass-weighted Deegan–Packel index. Theorems 1 and 3 admit analogous Holler–Packel and Responsibility lifts, but I would not count them as additional anchors.
The expected complexity is mixed. If AXp sets for the \(\tau\) types are explicitly available, aggregation is plainly Class A: compute each type’s contribution and sum \(\tau\) rational vectors. If AXp’s can be generated with polynomial delay for a model family, the continuous wrapper remains tractable in the high-multiplicity parameters. But continuity does not cure intrinsic explanation hardness. Concentrating all mass on one type, \(\mu_t=1\), recovers the original single-instance explanation problem. For succinct classifiers where exact AXp computation is hard, that hardness transfers; the population layer is not itself a new source of difficulty. Further questions would be whether scores can be computed without enumerating every AXp, whether additive approximation is possible when \(\tau\) is large, and which classifier representations admit efficient separation or enumeration.
My weakest point is decisive: this is a sensible continuous-population extension, but it is not a continuous mirror of a named computational theorem in the paper. It also changes the paper’s local task into a population audit unless the per-type outputs are retained. A strict referee could therefore reject it as outside the requested anchor discipline. The positive case survives only in the weaker sense that the paper’s axiomatic aggregator naturally defines a high-multiplicity population problem; it does not provide the computational result that ChoCo normally wants to continuize.
The proponent’s opening concession is the decisive negative verdict under ChoCo’s rules: this paper supplies no eligible computational anchor. Theorems 1–3 are uniqueness characterizations of aggregators; Proposition 4 is an axiomatic impossibility result. None states a complexity classification or an algorithmic theorem. “Computing abductive explanations is generally intractable” is an uncategorized claim attributed to earlier work, not a result proved here, and the experiments are empirical comparisons. There is therefore no named computational result whose continuous mirror this paper can claim.
The proposed Population Deegan–Packel problem is nevertheless a legitimate high-multiplicity construction. I would not object that repeated applicant profiles are impossible: fixed classifiers applied to recurring administrative profiles are plausible. The problem is that it is a new population statistic, not a population version of Theorem 2. The theorem concerns a map \((x,f)\mapsto\beta(x,f)\) for one instance. It says nothing about maps \((\mu,f)\mapsto z\). The expected-value formula is obtained by adding a new mass-linearity axiom, and under that axiom the formula is essentially stipulated. The paper provides no population characterization or computational result for it.
More fundamentally, the population does not participate in the object being explained. Changing \(\mu\) changes only how local explanations are weighted; it does not change the classifier, any AXp, any prediction, or any collective decision. Returning every per-type vector reduces the construction to a batch of the paper’s original local tasks; returning only \(z\) turns it into a lossy descriptive audit. Either choice may motivate a new XAI study, but neither is a continuous mirror of a computational result in this paper.
The strongest rescue would be to define a genuine computational problem for a succinct classifier: exact or approximate population expectations, score distributions, or top-\(k\) feature probabilities without enumerating all AXp’s. That could be worthwhile, but it requires choosing a model representation, encoding, approximation notion, and computational objective absent from the paper. It is a new counting/enumeration programme built around the paper’s objects, not a continuization of its theorems. Adding interventions such as reweighting profiles or repairing a classifier would make the problem more consequential, but would likewise introduce an external optimization problem rather than mirror this paper.
The Holler–Packel and Responsibility theorems suffer exactly the same defect, and Proposition 4 offers no computational anchor either. Thus I would reject this paper under the programme’s actual screening criterion. I would not honestly defend the stronger literal claim that no worthwhile population-XAI problem could ever be built from it: the applicant-profile example is a credible seed. The negative case is strong as a verdict about this paper’s eligibility, but weak as a universal claim about every possible re-modelling.
The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.