| paper | Combining Voting and Abstract Argumentation to Understand Online Discussions |
| authors | — |
| venue | AAMAS 2024 |
| filed under | multiwinner · multiwinner |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1) |
| judge confidence | high |
| authors would recognise it | yes |
Proposition 9
statement extracted from the paper’s text layer
Given an explicitly represented argumentation framework \(F=(\mathrm{Arg},\mathrm{Att})\), its preferred-extension catalogue \(E=\mathrm{prf}(F)\), a rational distribution \(\mu\) over nonempty approval types \(A\subseteq\mathrm{Arg}\), and \(k\), find \(\Omega\subseteq E\) with \(|\Omega|\le k\) maximizing \(U_\mu(\Omega)=\sum_A\mu_A\max_{\pi\in\Omega}|\pi\cap A|/|A|\); alternatively, decide whether an outcome core-represents every positive-mass type perfectly.
Types are nonempty approval ballots \(A\); mass is \(\mu_A\); decisions are sets of at most \(k\) preferred extensions \(\Omega\); the objective is mass-weighted average representation, or perfect core coverage.
The continuous population contributes only rational weights to a finite preferred-extension selection problem, and the hardness can already occur with one voter; this limits the mirror's population-specific novelty but does not defeat its validity.
fatal: False
The mirror covers the weighted forms of Propositions 1, 5, 7, 8, 9 and Theorem 4, while leaving the axioms, simulations, argument-mining pipeline, and non-computational semantic results untouched.
There is a credible continuous mirror here, and it is almost already latent in the paper’s notation. The mirror is not a fractional argumentation semantics: the attack graph and preferred extensions remain discrete. What becomes continuous is the electorate choosing among them.
Fix an argumentation framework \(F=(\mathrm{Arg},\mathrm{Att})\), with \(q=|\mathrm{Arg}|\), and let \(E=\mathrm{prf}(F)\). A voter type is a nonempty approval ballot \(A\subseteq\mathrm{Arg}\). A continuous society is a rational distribution \(\mu\) over approval types: \(\mu_A\) is the fraction of participants approving exactly \(A\). Only positive-mass types need be listed. A selected outcome is \(\Omega\subseteq E\), with \(|\Omega|\le k\). Thus the population is continuous, while viewpoints remain ordinary preferred extensions.
This is a particularly plausible high-multiplicity regime for a large, moderated consultation about a fixed policy proposal. Once the discussion has been formalized into a finite argument set, thousands or millions of participants may fall into a few dozen or hundred recurring approval patterns generated by political blocs, campaigns, or shared viewpoints. The paper’s own Example 2 already presents the electorate in grouped form, such as \(33\times\{p_1\}\) and \(31\times\{p_1,p_2,p_3\}\). Replacing those counts by their proportions is therefore a natural high-multiplicity representation, not a change of problem.
My lead is the following continuous problem, anchored in Proposition 9, with Proposition 7 providing the simpler unconditional hardness anchor. Both results are proved in this paper.
Call it Continuous Utilitarian Viewpoint Selection. Its input is \(F\), an explicitly listed preferred-extension catalogue \(E=\mathrm{prf}(F)\), a rational distribution \(\mu\) over approval types, and \(k\). For \(A\) and \(\Omega\), define
\[ r_A(\Omega)=\max_{\pi\in\Omega}\frac{|\pi\cap A|}{|A|}. \]
The task is to return a set \(\Omega\subseteq E\), \(|\Omega|\le k\), maximizing
\[ U_\mu(\Omega)=\sum_A \mu_A r_A(\Omega). \]
This is exactly the paper’s Utilitarian rule after continuization: for a discrete profile with \(n_A\) voters of type \(A\), set \(\mu_A=n_A/n\). Then
\[ U_\mu(\Omega)=\frac{1}{n}\sum_{i\in N}\mathrm{rep}_i(\Omega), \]
so the optimal outcomes are identical.
The expected classification is Class B: hardness transfers. Proposition 9 states that, assuming \(P\ne NP\), no algorithm can optimize the Utilitarian rule in time polynomial in \(m=|\mathrm{prf}(F)|\), even when the preferred extensions are supplied as part of the input. Proposition 7 gives the corresponding NP-hardness statement for optimizing an OWA rule, of which Utilitarian is the vector \((1,\ldots,1)\) case. A polynomial algorithm for the continuous problem would solve every discrete instance by setting \(\mu_A=n_A/n\). Moreover, duplicating every voter any number of times preserves \(\mu\), so the same computational question applies to arbitrarily large electorates without increasing the number of types.
The significance is not that continuization makes this problem easy. The hardness lives in selecting a small collection of mutually structured preferred extensions, and therefore survives the population compression. That is exactly the kind of Class B boundary the ChoCo programme wants to identify. The natural follow-up questions are whether the weighted problem is fixed-parameter tractable in \(k\), \(q\), or the number of approval types; whether the explicit-extension version admits approximation guarantees; and whether one can combine a greedy \(1-1/e\)-type argument with an efficient procedure for generating preferred extensions.
A second, independent mirror is Continuous Core-\(1\)-Representability, anchored in Theorem 4, proved in this paper. For each approval type \(A\), let
\[ \eta_A=\max_{\pi\in E}|\pi\cap A|, \]
and define its core representation by
\[ r_A^{\mathrm c}(\Omega)= \max_{\pi\in\Omega} \frac{|\pi\cap A|}{\eta_A}, \]
with the paper’s convention that the score is \(1\) when \(\eta_A=0\). The instance consists of \(F\), a rational distribution \(\mu\), and \(k\). The question is whether there exists \(\Omega\subseteq E\), \(|\Omega|\le k\), such that
\[ \sum_A \mu_A\, \mathbf 1\!\left[r_A^{\mathrm c}(\Omega)=1\right]=1. \]
In words: can at most \(k\) preferred extensions core-represent society almost everywhere? Since all members of a type have the same ballot and therefore the same score, this is precisely the paper’s requirement that every voter be perfectly core-represented.
Theorem 4 proves that the discrete problem is \(\Theta_2^P\)-complete, with hardness even for \(n=2\). The continuous version consequently remains \(\Theta_2^P\)-hard with only two positive-mass types: assign mass \(1/2\) to each of the two ballots in the reduction. Replicating those two types to an electorate of any size leaves the instance unchanged at the type level. This is especially strong evidence for the mirror: the problem remains hard in a genuinely high-multiplicity society consisting of two large blocs.
The same question also inherits the paper’s positive parameterized result, Proposition 5, proved here: \(1\)-Representability and \(1\)-Core-Representability are FPT in the number \(q\) of arguments. The enumeration of preferred extensions and candidate outcomes has dependence only on \(q\), with polynomial dependence on the number of listed types and on the encoding length of their masses. Thus the continuous formulation exposes a useful boundary: large population size disappears from the running time, while argumentation structure remains the source of combinatorial difficulty.
This mirror covers the paper’s computational results on representability and viewpoint selection—Propositions 1, 5, 7, 8, 9, and Theorem 4 in their corresponding weighted forms. It does not claim to continuize the paper’s axioms, simulations, or argument-mining pipeline. Those concern different objects and are outside the population-continuization scope.
The weakest point is that exact approval-type repetition may be less convincing on an open-ended online platform than in a fixed public consultation. If every participant approves a unique set of arguments, then \(\tau\) is close to \(n\), and the continuous representation offers little compression. The case therefore depends on the moderated, fixed-agenda regime: a large electorate, a stable finite argument set, and recurring approval patterns. That is narrower than “all online discussions,” but it is a natural regime explicitly compatible with the paper’s own grouped ballots and formalized-discussion model.
The strongest case against is that both proposed mirrors put continuity in the wrong place. Once \(F\) and \(E=\mathrm{prf}(F)\) are fixed, a voter contributes only through an approval set \(A\). For every outcome \(\Omega\), the paper’s objectives have the form
\[ \sum_A \mu_A v_A(\Omega), \]
where \(v_A(\Omega)\) is already determined by the finite argumentation instance. Replacing voter counts by rational coefficients is therefore a high-multiplicity encoding, but it introduces no population-level operation: no mass can be moved, split, targeted, or otherwise acted upon. All combinatorial structure remains in \(F\), its preferred extensions, and the choice of \(\Omega\).
This is most damaging to the Utilitarian anchor based on Propositions 7 and 9. The proposed objective
\[ U_\mu(\Omega)=\sum_A \mu_A \max_{\pi\in\Omega}\frac{|\pi\cap A|}{|A|} \]
is certainly well-defined, and the discrete-to-continuous correspondence is correct. But the same observation applies to every stronger version: weighted OWA rules, mass-weighted coverage, approximation objectives, or arbitrary rational type weights. They are weighted finite extension-selection problems. Proposition 1 makes the point especially sharply: NP-hardness already holds with one voter. Propositions 7, 8, and 9 derive hardness from selecting and covering preferred extensions, not from population multiplicity. The continuous formulation would faithfully preserve that hardness, but the population limit has contributed no structure of its own.
That is not, by itself, a valid objection under ChoCo’s rules: a dull Class B answer is still a legitimate result. It does, however, make this a weak candidate for the programme’s central purpose. The proposed mirror does not expose a continuous optimization problem; it merely replaces repeated rows in an objective function by weights.
Theorem 4 gives a more substantive degeneration. In the proposed continuous \(1\)-Core-Representability problem,
\[ \sum_A\mu_A\mathbf 1[r_A^{\mathrm c}(\Omega)=1]=1, \]
the numerical values of the positive masses are irrelevant. The condition is equivalent to requiring perfect core-representation for every type in the support. Thus a society with masses \(0.999\) and \(0.001\) is computationally identical to one with masses \(0.5\) and \(0.5\). The “continuum” disappears and only support membership remains. The theorem’s hardness with two voters consequently does not demonstrate a meaningful continuous population phenomenon; it demonstrates that the underlying argumentation problem is hard even before population size matters.
One could repair this by asking for at least \(1-\varepsilon\) represented mass, or by maximizing represented mass. But then the core anchor becomes a weighted coverage problem over preferred extensions, essentially another version of the Utilitarian or MaxCov formulation. The repair is sensible, yet it abandons the distinctive exact-core question that Theorem 4 supports. Proposition 5 has the same limitation: its FPT algorithm is parameterized by the number of arguments and enumerates extensions; dependence on the electorate is merely polynomial bookkeeping over listed ballot types. It is an argumentation result with weighted inputs, not a continuization result.
The proposed high-multiplicity story is nevertheless plausible. A fixed, moderated consultation can have millions of participants distributed among recurring approval patterns, and Example 2 already groups voters by ballot. That defeats any claim that multiplicity is impossible, that voter identity is essential, or that the mirror is formally invalid. Nor can existing high-multiplicity work be used against it.
Accordingly, the honest negative case is narrow: these anchors are valid continuous reformulations, and the proponent is right that they yield Class B boundary results. I would oppose them only under a stricter conception of “worthwhile mirror,” requiring the population continuum to affect the problem beyond supplying weights. Under the programme’s stated definition, that stricter conception is not justified. The universal claim that no worthwhile mirror exists is therefore not supportable; at most, the paper offers a modest and largely population-insensitive continuous mirror.
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.