| paper | d4-conf-ijcai-2019-0603 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | medium |
| authors would recognise it | yes |
Both anchors are computational results, but the proposed cohort objective is exactly the paper’s expected time-to-rank objective under \u\mathcal D\). Recasting that distribution as actual cohort composition changes the interpretation, while the paper already gives the four-ranking hardness result and the uniform-distribution PTAS; this is a \(b\text{-already}\) collision.
fails bit b-already — the paper is already native to the continuum
The paper already optimizes expected Kendall distance over a distribution of target rankings, including the uniform four-ranking case and uniform finite lists; the cohort reading adds no new computational instance. The proponent identifies this equivalence but does not establish a computational distinction between personalized uncertainty and shared-cohort composition.
fatal: True
The cohort reading maps directly to the paper’s uniform four-ranking hardness result and its PTAS for uniform ranking distributions; it adds no result beyond those distributional cases.
The strongest mirror is a shared recommendation for a large cohort. Imagine an organization asking many people to rank the same meeting slots, and showing everyone the same initial ordering. Each type \(t\) is a complete user type: a target ranking \(\pi_t\), with the same interface and linear drag-time model as in the paper. Its mass \(\mu_t\) is the fraction of users with that type. The platform chooses one strict ranking \(\sigma\), minimizing average time-to-rank:
\[
\min_{\sigma\in\mathcal L(A)}\ \sum_{t\in T}\mu_t\,\mathrm{dkt}(\sigma,\pi_t).
\]
The recommendation stays a discrete ranking; it is the population that becomes continuous. A natural regime is a large organization with many users but few preference clusters—for example, \(N\) users and four common schedules, so \(N\gg\tau\). Replicating each cohort changes neither the masses nor the best recommendation.
My lead anchor is Theorem 3(c), proved in this paper. It says that \(\Rec\) is NP-complete when the weight function is linear and \(\D\) is uniform over \(n=4\) linear orders. The paper proves this by reducing from Kemeny ranking with four voters, whose NP-completeness it cites from elsewhere. In the cohort mirror, the corresponding question is: given four ranking entries, each with mass \(1/4\), and a threshold \(\delta\), is there a shared recommendation \(\sigma\) whose average time-to-rank is at most \(\delta\)? The answer is yes exactly when the four-order Kemeny instance has total Kendall distance at most \(4\delta\), since the paper’s linear cost equals Kendall distance. A ranking is the solution.
I expect this anchor’s exact hardness to transfer (Class B). The reduction’s combinatorics lie in the number of alternatives, not in having many voters. A four-entry instance can represent cohorts of \(q\) identical users each, for any \(q\); the population then has \(4q\) people but still only four types. The continuous formulation preserves the objective after normalization.
A useful second anchor is Theorem 4(c), proved in this paper using a cited result. It says that \(\Rec\) has a PTAS for linear weights and a uniform distribution. The paper establishes the recommendation result by reducing it to Weighted Feedback Arc Set in Tournaments and applying the PTAS of Kenyon et al. In the mirror, given a list of \(n\) ranking entries with equal mass \(1/n\), return a shared ranking whose average time-to-rank is at most \(1+\varepsilon\) times optimal. The paper gives running time \(m^{2^{\widetilde O(1/\varepsilon)}}\), apart from the work to read and aggregate the input. This predicts approximation tractability on the uniform-list subclass, even though Theorem 3(c) rules out exact polynomial-time optimization unless \(\mathrm{P}=\mathrm{NP}\). The result does not by itself cover arbitrary binary-rational masses.
These are two guarantees for the same underlying cohort recommendation problem, on stated input subclasses: exact hardness at four equal-mass ranking entries, and a PTAS for uniform finite lists. The mirror covers only time-to-rank recommendation under linear move costs. It does not turn rankings into fractional outcomes or study a probability model over one user’s preferences. The authors should recognize the shared-cohort version as their problem: the recommendation decision and time objective are unchanged, and the paper itself motivates ranking tasks involving participants.
The weak point is also the main interpretive risk: the paper already represents a user’s unknown target ranking by a distribution \(\D\) and minimizes expected cost over it. Mathematically, that objective is the same as averaging over a population. The case for a mirror depends on the paper’s stated interpretation of \(\D\) as uncertainty about one user, while \(\mu\) records the actual composition of a large cohort receiving one common recommendation. If the authors already regard \(\D\) as an aggregate population distribution, the mirror may be a restatement. If they mean personalized preference uncertainty, the cohort question is distinct and fits the population-continuization scope. A natural next question is whether the PTAS extends to arbitrary rational cohort masses, with complexity stated in \(m\), support size \(\tau\), and mass bit length \(L\).
The difficulty for the cohort mirror is that the paper has already written down its objective. Given types that differ only by target ranking, let \(\mu_t\) be each type’s population share. The cohort’s average time is exactly \(\mathbb E_{\tau\sim\mu}[\mathrm{dkt}(\sigma,\tau)]\), the objective in \(\Rec\). The chosen ranking, feasible solutions, and costs are unchanged; only the interpretation of \(\D\) changes from uncertainty about one user to the composition of a group.
That makes the hardness anchor an exact overlap: four equally weighted ranking types are the paper’s uniform \(n=4\) case. Replicating each type \(q\) times changes neither the normalized input nor the objective. The hardness is already proved for that recommendation problem, so the cohort reading adds no computational question.
The PTAS anchor overlaps just as directly. The paper gives a PTAS for uniform distributions over rankings, precisely the equal-mass cohort version. Even the stronger version with arbitrary rational masses looks immediate from its proof: for each pair \(a_i,a_j\), sum the mass of types that prefer \(a_j\) to \(a_i\). Those pairwise weights define the same weighted feedback arc set instance used in the paper’s PTAS reduction. The four-type hardness case remains a subcase. Thus allowing support size \(\tau\) and mass bit length \(L\) does not appear to open a substantive new problem.
The most plausible way to escape this overlap is to give types different drag-time functions as well as different rankings. That is a sensible extension, especially given the paper’s user-behavior experiments, but it changes the optimization model: the paper’s computational results use one common weight function. I can’t honestly rule out that heterogeneous-cost cohort problem being worth studying. It is a weaker mirror of the named results, though, and it does not rescue the proposed ranking-distribution anchors.
So the negative case rests on collision, not on a claim that large cohorts or mass are meaningless. For the stated anchors, the paper already optimizes the cohort-average objective and gives the cited hardness and approximation guarantees. My case is weaker if the authors treat the shift from predictive uncertainty to an actual shared cohort as a research question in its own right—but on the mathematical formulation given here, it looks like a relabeling of \(\D\).
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.