| paper | d4-conf-aaai-2010-0074 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
The paper contains computational formulations, procedures, and experiments, but no numbered theorem-like result asserting a computational claim. Its numbered Definition 1 and unnumbered formulations do not qualify as anchors. The proposed shared-product or cohort-query model adds collective coupling absent from the paper’s individual-user recommendation problem.
fails bit a — no named computational result to mirror
No qualifying named computational result is covered; the formulations and experiments concern individual-user elicitation, while the shared-population repair changes the decision problem.
I can’t make a compliant positive case with the required anchors: the paper contains no numbered Theorem, Lemma, Corollary, or Proposition asserting a complexity result. Its closest candidate is “Computing Regret: Conjunctive Concepts”: equations (4)–(6) formulate minimax-regret recommendation as an integer program, and equations (7)–(15) formulate the max-regret subproblem. The authors use these formulations in their computational experiments, but they are not a numbered result of the kind the anchor rule requires. Definition 1 defines the regret problem; it does not classify its complexity.
A population version is imaginable: a recommender serves many customers drawn from finitely many preference types, with mass \(\mu_t\) for each type, and chooses a shared product to minimize population regret. But if types specify the actual utility and concept, this becomes aggregate utility optimization, not the paper’s minimax decision under one user’s utility and concept uncertainty. If each type gets its own recommendation, the decisions separate by type. Either way, the population model changes the underlying question, and I would not claim the authors would accept it as a continuous mirror of a named result here.
So there is no defensible per-anchor continuous problem to hand to a prover, and no tractable, hardness-transfer, or continuum-specific classification to predict. The paper does compute with MIP formulations, but treating those as qualifying anchors would stretch the requirement.
The strongest case against a direct mirror is that the paper has no numbered complexity result to mirror. Equations (4)–(15) give integer program formulations and a constraint-generation procedure; the experiments compare elicitation strategies. Those are real computational contributions, but the paper makes no complexity classification, approximation result, or parameterized claim about a population problem. Under the anchor rule, the proponent names no qualifying result.
The deeper mismatch is that the computational object is one user’s uncertain utility and subjective concept, used to choose that user’s product. The hard parts are the feasible product set, utility polytope, and concept version space. There is no electorate, shared allocation, or query budget coupling users. Repeating an identical user adds no new optimization: their mass scales regret without changing the best recommendation. Giving each type its own recommendation makes the decisions separable.
The strongest repair is a population that must share one product or default, with queries allocated across preference types to reduce population regret. That can be sensible for a fleet, workplace, or platform default. It also introduces the coupling that makes mass matter: a shared decision or limited query budget. Those choices create a collective recommendation or budgeted elicitation problem absent from the paper. If each type means users with exactly the same utility and concept, one representative’s answer can inform the whole cohort; if preferences vary independently, the support may grow with the number of users and offer little high-multiplicity compression.
That is a strong reason to reject the paper as a straightforward ChoCo target. It does not establish that no worthwhile continuous mirror exists in any scenario. A shared-default or cohort-query model is coherent enough to survive the negative case, though it would be a new population problem rather than a continuous version of a named complexity result in this paper. The honest negative is therefore strongest against direct mirroring and weak against the universal claim.
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.