| paper | d4-conf-aaai-2020-1030 |
| 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 describes an approximate LBP procedure and reports experiments, but it has no numbered theorem, lemma, corollary, or proposition asserting a computational result. That fails the required computational anchor, regardless of whether a cohort-level extension could be formulated.
fails bit a — no named computational result to mirror
No named computational result qualifies as an anchor; the LBP procedure and empirical comparisons are not anchored computational guarantees.
The paper has no qualifying named computational result to anchor a continuous mirror on. It states a prediction task in Equation (1), gives an approximate LBP learning procedure in Algorithm 1, and reports empirical performance in Table 2. It contains no numbered theorem, lemma, corollary, or proposition establishing a complexity result such as polynomial-time solvability, hardness, or fixed-parameter tractability.
Promoting Algorithm 1 into a tractability anchor would overstate what the paper proves: it uses approximate message passing on potentially cyclic graphs and gives no runtime or convergence guarantee. A population-level version of sentiment prediction might be worth formulating, but it would be a new model question, not a mirror of a named computational result in this paper.
The absence of a complexity theorem is accurate, but it is not the strongest objection. TSSM defines a prediction task in Equation (1), a named probabilistic model, and learning and inference procedures in Algorithm 1 and Equation (13). Those give a continuous version something to ask about, even though they do not establish a worst-case complexity result.
The more serious problem is that TSSM predicts sentiment for particular users in a particular temporal graph. Its factors depend on each user’s past labels, tweets and interests, and on the specific links and learned strengths between users. An exact type would therefore have to capture those features and enough of each user’s graph neighborhood to preserve the relevant dependencies. In a realistic network, that can leave nearly every user in a separate type: \(\tau\) grows with the population, so the proposed high-multiplicity compression disappears. Grouping users only by interest or topic profile would discard information the paper’s inference task uses.
The strongest repaired mirror would model exchangeable cohorts, provide a type-to-type influence kernel, and predict the future sentiment distribution within each cohort. That is a coherent computational question. But it changes the input from a specified graph with pair-specific influences to an aggregate network model, and changes the output from predictions for named users to cohort-level predictions. If exact links are retained, the type classes tend to collapse; if links are averaged, the resulting model is a new block-network or mean-field formulation. It may be a worthwhile extension, but the paper does not supply evidence that this is its natural high-multiplicity regime.
So I would not use “no numbered complexity result” as a knockdown argument, and I cannot honestly rule out every plausible cohort scenario. The negative case is narrower: for ChoCo, this paper is a poor mirror candidate because its computational contribution is tied to identity-specific graph inference, while the population-level alternative requires a substantial new network model and a different prediction target.
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.