| paper | d4-conf-atal-2023-0044 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | yes |
The paper defines preference prediction and norm inference, then reports empirical results, but states no numbered computational result of the required kind. The mass based prediction query is a plausible extension, yet it is not a continuous analogue of a named computational result in this paper.
fails bit a — no named computational result to mirror
The proposed extension would cover similar user selection, average preference prediction, and preference completion; it leaves confidence measures, norm thresholds, norm inference, and empirical validation outside the computation.
The paper has no valid named computational-result anchor under your rules. It contains no numbered theorem, lemma, corollary, or proposition establishing a complexity result. Definitions 7–10 specify similar users, prediction, and completion; Section 5 reports empirical validation. Those are useful foundations for a proposed mirror, but they are not results classifying a computational problem. So I cannot honestly present an anchor-based case or assign one continuous problem per anchor.
The strongest prospective mirror I see is a continuous version of the paper’s collaborative-filtering prediction. A type would be a user’s complete recorded partial-preference profile over a fixed menu \(X\), including which preferences are unknown; mass would be the fraction of a large assistant-user cohort with that profile. Given a target user’s partial profile \(q\), an item \(x\) whose preference is unknown, a separation function such as Def. 6’s cumulative separation, a radius \(\epsilon\), and a minimum cohort share \(\alpha\), the instance asks for the prediction obtained by averaging the preferences for \(x\) among types within \(\epsilon\) of \(q\), together with the nearest types needed to reach mass \(\alpha\). Include all types tied at the cutoff. The solution is the selected type mass and its mass-weighted average preference; applying the paper’s thresholds can then produce a norm.
A plausible regime is a large platform user base answering a standardized, bounded menu of privacy questions, with many users sharing exact partial profiles, so the number of users greatly exceeds the number of distinct types. This preserves the paper’s personalized nearest-user prediction and changes its minimum-neighbor count to a minimum population share. I would expect this computation to be tractable over a listed support, roughly \(O(\tau |X|+\tau\log\tau)\) for one query. The soft spot is substantial: the paper’s 825-question survey and varying missingness make repeated exact profiles uncertain, and the authors establish no complexity result for this operation. This is a plausible continuous analogue of its prediction method, but it does not meet the required anchor standard. It covers preference prediction, not norm inference as a computational result.
The case against begins with the required anchor: this paper states no numbered computational result to continuize. It defines preference approximation and norm inference, proposes a prediction function, and reports experiments. It gives no complexity classification, algorithmic theorem, or hardness result. The other analyst acknowledges this, so their proposed mirror is a new computational question inspired by the paper, not a continuous version of one of its results.
The strongest setting for that new question is plausible: a large platform with a fixed privacy questionnaire and many users sharing partial profiles. Given an explicit support of \(\tau\) profile types, the proposed prediction scans types with a known answer to \(x\), computes the cumulative separations, selects types within \(\epsilon\) and enough nearest mass to reach \(\alpha\), then takes a weighted average. Sorting handles the cutoff; thresholding converts the result to a norm. This is a straightforward weighted nearest-neighbor query, polynomial in \(\tau\), \(|X|\), and the input precision. Running it for every supported target type remains polynomial. It offers no computational bottleneck of the sort the programme is set up to chart.
There is also a mismatch between the paper’s data and the compression the mirror hopes to gain. A type defined by a user’s entire partial profile must include which questions they answered as well as their answers. The paper’s survey has 825 questions, with participants answering different random subsets of roughly 144–199. That creates a huge space of possible types; the reported cohort does not establish repeated profiles. Coarsening profiles could create more mass per type, but it changes the distances and predictions. A different, smaller common menu could make repetition plausible, yet the computation would still be the same simple weighted-neighbor operation.
A stronger proposed variant might optimize thresholds or predictors for low expected error across the cohort. That needs a distribution linking each observable partial profile to the user’s true, withheld preferences. A distribution over observed profiles alone cannot determine whether a prediction is accurate. Supplying the true preferences as part of each type gives the predictor information it is meant to infer; supplying labeled examples instead makes this a statistical-learning problem. Either way, the paper’s per-user norms remain individual outputs, with no collective choice or social outcome to compute.
So the best negative case is that there is no result in this paper for the programme to mirror, and its most natural continuous extension is a routine data query rather than a computational social choice problem. I would not claim the universal impossibility literally: a cohort-based prediction study could be well posed and useful to its authors. The honest weakness is that usefulness alone does not establish a worthwhile ChoCo mirror, and the paper supplies neither a computational result nor a nontrivial computational problem to carry over.
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.