| paper | d4-conf-ijcai-2015-0249 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | unclear |
The paper describes CBACE models, approximate inference, and empirical evaluations, but contains no numbered result asserting a computational property. The proposed population-weighted inference problem is coherent, yet it is a new, unanchored question rather than a continuous analogue of a qualifying result in this paper.
fails bit a — no named computational result to mirror
The proposed extension concerns CBACE and CBACE+Bias aggregation; the paper’s accuracy comparisons and community-learning evaluations do not provide a named computational result to mirror.
I find no qualifying anchor. The paper has no numbered theorem, lemma, corollary, or proposition asserting a complexity result such as polynomial time, NP-hardness, or fixed-parameter tractability. Equations (4)–(5) specify the CBACE model and posterior; the paper then evaluates approximate inference empirically. Its runtime report and accuracy improvements are not named computational results. So there are no anchors to quote, and no anchor-indexed continuous problems or complexity predictions to give.
The best positive lead I can make is an unanchored population version of CBACE aggregation. Imagine a large Wi-Fi sensing programme with many reporters but few recurring reporting profiles—for example, combinations of device and firmware, sensor mode, and task context. A type is a complete profile for the aggregation model; \(\mu_t\) is its fraction of reporters. Given a total population scale \(Q\), let \(D_t\) be the observed distribution of reporters’ full report histories \(h\) within type \(t\). Keep CBACE+Bias’s latent community, precision, and bias model, with a fixed number \(M\) of communities and the paper’s priors.
For item values \(z=(z_i)\) and community parameters \(\psi=(\pi,\tau,\beta)\), define a history’s likelihood by integrating out a reporter’s latent community \(m\), precision \(a\), and bias \(b\):
\[ p_\psi(h\mid z)=\sum_{m=1}^{M}\pi_m\int_0^\infty\int_{\mathbb R^n} \operatorname{LogN}(a\mid\tau_m,\alpha_0)\, \mathcal N(b\mid\beta_m,\Theta_0) \prod_{(i,x,\theta)\in h} \mathcal N(x\mid z_i+b,a\theta I)\,db\,da. \]
The continuous problem would be: given the rational masses \(\mu_t\), the finite-support history distributions \(D_t\), and the priors, compute the posterior means and covariances of the item values under
\[ p(z,\psi\mid\mu,D)\ \propto\ p_0(z,\psi)\exp\!\left( Q\sum_t\mu_t\sum_h D_t(h)\log p_\psi(h\mid z) \right). \]
This makes population mass the continuous object; the real-valued measurements and Bayesian uncertainty are already features of the paper, not the proposed continuization. The scenario is plausible if many reporters share a small set of complete, recurring profiles. For a fixed \(M\), approximate inference looks like a possible Class A direction, since the paper already uses variational message passing. But the paper gives no complexity guarantee for that method, and I would not classify exact posterior computation from these results.
The weak point is substantial: CBACE gives each user individual latent precision and bias. Grouping users who share only a device or context may therefore fail the programme’s complete-type test; grouping by their actual reliability could make the number of types large. The proposed mirror also extends the paper’s model to mass-weighted cohort data. It is a plausible question for the authors, but the paper does not establish it as a computational result. My positive case covers the aggregation model and its inference problem, not the reported accuracy gains as complexity claims.
There is no named computational result to anchor a mirror: the paper proposes CBACE models, uses approximate variational inference, and reports empirical accuracy and runtime. It gives no complexity theorem, algorithmic guarantee, or hardness result. So the proponent is right that their proposed population inference problem is unanchored: it is a new question inspired by the paper, not a continuous counterpart to one of its stated results.
The strongest version of that question is coherent: weight reporters’ full histories by population mass and ask for posterior estimates or community selection. But preserving the histories is the problem for the negative case. If a type includes a reporter’s complete history, real-valued reports and task-specific observations make most types distinct, so support can grow almost one-for-one with the number of reporters. If types instead capture recurring contexts such as device and firmware, the history distributions \(D_t\) still carry the individual evidence CBACE uses to infer reliability and bias. Coarsening those histories further risks discarding that evidence. When complete histories do repeat, the weighted likelihood is essentially a repeated-record dataset; that is a sensible high-multiplicity regime, but the paper gives no reason to expect it to be common or to unlock the continuous-optimization tools that motivate ChoCo. The paper’s central task is Bayesian estimation, too, rather than a clearly stated social-choice optimization problem such as control or bribery.
That makes the negative case credible but limited. It explains why this paper supplies no theorem-specific mirror and why the proposed mirror’s practical value is uncertain. It does not establish that weighted inference over repeated reporter types would be unworthy to ask: the authors could recognize it as a useful computational problem. I cannot honestly rule out every plausible scenario; the case against is weakest precisely at that unanchored proposal.
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.