| paper | d4-conf-ijcai-2024-0943 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | medium |
| authors would recognise it | yes |
Proposition 1
statement extracted from the paper’s text layer
Given a finite menu of alternatives in \([0,1]^n\), a fixed local aggregator \(\phi\), and a rational distribution \(\mu\) over report types \(t=(x_t,y_t,r_t)\), where \(r_t\) records strict preference or indifference, find a normalized Möbius vector \(m\) minimizing \(\sum_t \mu_t\ell_t(m)+\lambda\|m\|_1\), subject to monotonicity of \(v(S)=\sum_{R\subseteq S}m(R)\), optionally with supermodularity. Here \(\ell_t(m)=[\delta-\langle m,\phi(x_t)-\phi(y_t)\rangle]_+\) for strict preference and \(\ell_t(m)=[|\langle m,\phi(x_t)-\phi(y_t)\rangle|-\delta]_+\) for indifference.
A panel whose members each contribute a comparison report from a fixed menu and share a common latent capacity. The mass \(\mu_t\) of each report type weights its hinge loss; one sparse normalized capacity \(m\) is fitted to the panel, with the paper’s monotonicity and optional supermodularity constraints.
The stated population objective is static and treats respondent mass as weights on reports, while Proposition 1 concerns ordered online updates; the proponent gestures at mass arriving over time but does not specify that update or its guarantee.
fatal: False
The mirror covers constrained capacity fitting and the algorithmic direction of Proposition 1. It leaves aside the paper’s individual decision-maker setting, its numerical evaluations, and its experiments with particular Choquet models.
My lead, and the paper’s only numbered computational anchor, is Proposition 1. It is proved in the supplementary material. For online capacity learning with linear constraints \(Bm\le 0\), the proposition says that replacing the augmented Lagrangian with its linearization makes the ADMM–RDA updates closed form. In particular, the primal update soft-thresholds a gradient adjusted by the constraint multipliers,
\[
\bar g_t^\mu=\bar g_t-B^\top(\bar\mu_t-\rho(B\bar m_t-\bar z_t)),\qquad
m_{t+1}=-\frac{\sqrt t}{\gamma}[|\bar g_t^\mu|-\lambda]_+\operatorname{sign}(\bar g_t^\mu),
\]
and the split-variable and multiplier updates are
\[
z_{t+1}=-[\mu_t/\rho-Bm_{t+1}]_+,\qquad
\mu_{t+1}=\mu_t-\rho(Bm_{t+1}-z_{t+1}).
\]
Here the constraints can encode monotonicity and supermodularity. The paper uses this algorithm in its numerical experiments. This is an algorithmic result, not a named polynomial-time or hardness classification.
A plausible population mirror is Population Capacity Fitting\(_\infty\). Consider a large panel whose members compare alternatives described by the same fixed \(n\) criteria. A platform or committee wants one sparse capacity-based model of the panel’s collective preferences. Suppose each member answers one comparison from a fixed finite menu. A type \(t\) records the comparison \((x_t,y_t)\), the reported strict preference or indifference, and any other respondent information the fitting objective uses. Members of a type make identical contributions to the problem. The society is a distribution \(\mu\) over its \(\tau\) observed types, with rational masses summing to one.
Fix the paper’s local aggregators \(\phi\), such as \(\phi_S(x)=\min_{i\in S}x_i\) for a Choquet model. The decision variable is a Möbius vector \(m\), defining \(F_m(x)=\langle m,\phi(x)\rangle\). For a strict report \(x_t\succ y_t\), let the loss be
\[
\ell_t(m)=[\delta-\langle m,\phi(x_t)-\phi(y_t)\rangle]_+;
\]
for indifference, let it be
\[
\ell_t(m)=[|\langle m,\phi(x_t)-\phi(y_t)\rangle|-\delta]_+.
\]
The problem is to find a sparse normalized capacity minimizing
\[
\sum_t\mu_t\ell_t(m)+\lambda\|m\|_1,
\]
subject to \(v(\varnothing)=0\), \(v(N)=1\), and monotonicity constraints on \(v(S)=\sum_{R\subseteq S}m(R)\). Supermodularity can be required as an additional linear constraint. A solution is an optimal capacity, hence a scoring model for the panel’s alternatives; an online version would update that model as type mass arrives.
This regime could describe millions of users comparing products from a fixed catalogue of attributes, with many users answering the same query and giving the same response. The number of people can be far larger than the number of distinct query-response types. Since the paper’s objective is preference-example loss, the population-weighted objective preserves what is being learned: the capacity-based preference model. In a discrete panel with counts \(n_t\), setting \(\mu_t=n_t/\sum_s n_s\) gives the same average fitting objective. Weighting types by mass is the natural high-multiplicity representation. It also makes the model relevant to a collective decision maker, rather than only to the individual DM emphasized in the paper.
I expect Class A when the number of criteria is fixed or modest. The weighted hinge-loss problem with these constraints is a linear program after introducing slack variables, so its size is polynomial in the type support and the explicit capacity dimension \(2^n\). That supports solving the problem independently of the number of individual respondents. The paper’s updates similarly suggest processing weighted feedback without replaying every person’s record, though the weighted update itself would need to be derived and analyzed. For variable \(n\), the capacity representation and constraint system are exponential; Proposition 1 does not establish a polynomial-time algorithm in \(n\), and the paper gives no hardness result to classify that case. Further questions include whether an implicit representation can avoid the \(2^n\) dependence, and what regret guarantee holds when incoming observations carry rational mass.
The mirror covers the paper’s constrained online learning result and its capacity-fitting task; it does not continuize criteria, outcomes, or time as its central move. Its weakest point is that the paper learns one DM’s preferences, while this mirror fits a collective model from many respondents. The authors could reasonably object if they regard their task as strictly individual. The case survives only where a shared group-level capacity is an accepted target: there the same pairwise learning problem and loss are retained, with the evidence supplied by a population.
Proposition 1 is the paper’s only plausible computational anchor. It gives a real algorithmic result, but its object is an ordered stream of comparisons used to learn one decision maker’s capacity. Those comparisons are data examples, not agents in a society.
The strongest mirror would pool a large panel answering questions from a fixed menu, assume everyone shares one capacity, and give each comparison type its population mass. The resulting expected hinge loss with sparsity and capacity constraints is a coherent population-weighted fitting problem. The authors might recognize it as a collective extension of their learning task.
But the static population version no longer captures Proposition 1’s online contribution: it discards the order in which evidence arrives. To preserve online learning, give each round a distribution over feedback types and aggregate their subgradients. That produces a weighted version of the paper’s update, with the same capacity constraints and essentially the same analysis. The mass then weights training examples; it does not describe a society whose collective choice or strategic behavior is being computed. Compressing repeated reports may help with a large dataset, but it does not turn the online learning question into a distinct population problem.
The common-capacity assumption is also doing substantial work. If respondents have different underlying preferences, pooled reports fit an average model rather than the individual model the paper studies. Giving each person a separate capacity would model heterogeneity, but that is a different learning problem, with a different target and no anchor in this paper’s result.
So the negative case is strongest as a case against priority for ChoCo: the plausible mirror is a weighted-data extension of online preference learning, not a computational-social-choice problem about a continuous society. It is not airtight. A shared-capacity panel is a plausible scenario, and I cannot honestly rule out that the authors would recognize it as theirs. The case against a worthwhile mirror is therefore limited to its weak fit and likely small incremental contribution, not a claim that no such scenario exists.
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.