acceptodds
Under review as a conference paper at ICLR 2027

GeoRARE: Certified Rarity Profiles for Structural Causal Effects

Abstract

Rare treatment–context regimes create two distinct problems: deciding how much they matter and determining what the data can establish about them. GeoRARE defines a two-dimensional rarity-indexed structural causal profile that separately indexes treatment and context rarity. Sampling allocation governs information; reference weights govern the scientific question. One law-confidence event certifies the entire profile and its signed contrasts, with width controlled by unresolved reference mass. A finite sample upper bound and a rare-cell lower bound identify the same local information scale, up to constants and logarithmic terms. Consequently, profile statements can be certified before all individual structural signatures are resolved. Categorical certification experiments, a stochastic oscillator, signed profile reversals, and an ACIC-covariate benchmark examine this distinction and its extension under measured confounding. The same rarity family also provides a secondary targeting interpretation as an interpretable training-priority family under finite-sample, regularization, and capacity trade-offs; the experiments show no universal predictive advantage. The contribution is a rarity-indexed structural causal profile whose entire surface and signed contrasts can be certified simultaneously, together with upper and lower results that characterize its information limits.

Primary area:
causal reasoning
Keywords:
Structural Causal Models, Causal Inference, Rare Events / Rare Regimes, High-Dimensional / Rare-Cell Bounds

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_CbFfOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper appears to consider topological data analysis. Building on Faghihi and Saki (2026), the paper considers a summary of the difference in outcome distributions under alternative treatments. Let uazu_{az} be the conditional distribution (conditional on Z=zZ=z) of the potential outcome under treatment level aa, and ubzu_{bz} be the conditional distribution under treatment level bb. Consider some functional Ψ\Psi (the paper never explains exactly what Ψ\Psi is other than providing one example), then Ψ(uaz)−Ψ(ubz)\Psi(u_{az})-\Psi(u_{bz}) provides a causal contrast, in the sense that it measures the differences in features of the conditional potential outcome distributions between the two treatment levels. Faghihi and Saki (2026) consider estimation of the norm of Ψ(uaz)−Ψ(ubz)\Psi(u_{az})-\Psi(u_{bz}), the present work takes a weighted average of this norm over different levels of aa, bb, and zz to get a scalar summary, and chooses the weights to emphasize or de-emphasize rare values of zz, aa, and bb. The paper then provides `certification' of this object (i.e., construction of confidence sets and/or hypothesis testing). Some numerical experiments are carried out.

Critical strengths

Topological data-analysis is an interesting and a useful topic of study. Providing confidence sets for objects of interest seems a potentially useful innovation relative to existing work.

Critical weaknesses

The paper is poorly explained and dense with what seemed to me to be obscure/non-standard technical jargon. For example, the second sentence in the paper "Increased fluctuations around an equilibrium and the appearance of a low occupancy gap may both change an outcome distribution, while only the latter changes a prespecified connectivity signature." What equilibrium? Increased fluctuations of what? What is a "low occupancy gap"? What is a "connectivity signature" in this context? These terms were unfamiliar to me. Some of them seem to be from work on topological data analysis, particularly Faghihi and Saki (2026) on which the current work builds heavily. The latter paper is (at lest in my reading) much clearer than the present work about its goals and the meanings of the notation and technical terms it uses.

In my reading, one of the key contributions of the paper is to consider a weighted (over treatment levels and covariate levels) version of the causal contrast in Faghihi and Saki (2026). This seems to a fairly minor innovation. `Certification' (which here refers to construction of confidence sets/statistical hypothesis testing) seems like it may be a more important contribution, but I found the construction and discussion here difficult to understand, for example, the sentence "Weighted sums give LΘ(d,r)L_{\Theta}(d, r) and UΘ(d,r)U_{\Theta}(d, r)" was insufficient for me to understand what these objects were or why it is useful to construct a high-probability upper bound on their difference.

Questions for authors

What are LΘ(d,r)L_{\Theta}(d, r) and UΘ(d,r)U_{\Theta}(d, r)? What does the certification actually allow one to conclude? Could you explain what your methods achieve relative to the prior literature in the context of a realistic empirical example?

Reviewer_7JfUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces GeoRARE, a new framework for causal inference designed to handle situations in which specific treatment-context combinations are extremely rare. It establishes a two-dimensional "rarity-indexed structural causal profile" that separately tracks treatment rarity and context rarity.

Critical strengths
  1. Interesting topic on causal inference where complex treatments are rare in reality.
Critical weaknesses

see Qs.

Questions for authors
  • The first three paragraphs of the introduction section of this article lack relevant reference support. It is recommended that the author supplement relevant citations in the first three paragraphs
  • Equation (3) involves optimization solution, and it is mentioned in the text that it can be precisely enumerated when the number of vertices is small; In general, only exponential complexity can be solved. If exponential complexity is not the optimal solution, how much error does it have compared to reality? Is there a theoretical error upper bound guarantee? And what setting does the threshold c rely on?
  • Experimental evidence has shown that causal adjustment and unclear information are important parts of graph interpretation that cannot be ignored. So, which is more important for graph interpretation, causal adjustment or unclear information? Has the author quantified it? If quantifiable, which part is the dominant source; If direct comparison is not possible, can you explain the reasons behind it?
1 comment

Thank you for your response. The paper could benefit from further refinement.

Reviewer_WQY7OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper introduces GeoRARE, a rarity-indexed causal profile that separates scientific reference weights from the sampling allocation that determines precision. Each categorical potential-outcome law is mapped to a finite structural signature, such as the Betti numbers of a thresholded occupancy complex. Simultaneous confidence sets for the cell laws produce bands for the profile and for signed contrasts chosen after inspecting the bands. Theorem 1 bounds band width through unresolved reference mass, with cellwise information governed by sample size, sampling probability, and squared decision radius. Theorem 3 gives a rare-cell lower bound at the same scale up to logarithmic factors. The experiments verify the predicted certification transition and demonstrate a signed profile reversal that can be certified before all cell signatures are resolved. A measured-confounding extension shows that valid certification requires standardized causal laws and appropriate score uncertainty. The paper reports unresolved cases, clarifying the information limits of the proposed profile.

Critical strengths
  • Clear separation of target and information. Equation (2) defines the estimand through the reference weights, and Theorem 1 makes the sampling probabilities govern certification.
  • Theorem 1 and Proposition 4 cover the rarity surface, data-selected degrees, and signed contrasts on one confidence event.
  • The experiments include unresolved certificates, invalid confounding adjustments, and cases where the targeting method underperforms. These negative results help delimit the claim.
Critical weaknesses
  • Limited methodological novelty. The construction combines escort weights, simultaneous confidence-set projection, and finite structural signatures; the experiments omit a comparison with direct projection for the same prespecified target.
  • Restricted scope. The theory assumes a fixed finite categorical geometry, threshold, and reference masses. Computation poses a separate limitation because exact signature enumeration is exponential in general. Theorem 3 supplies a path-specific lower bound, leaving general minimax behavior open.
  • Most evidence is synthetic or semi-synthetic. The sci-Plex study is descriptive at the finite-panel level; population-level biological identification remains open.
Questions for authors
  1. Using the same laws, confidence budget, and structural descriptor, what additional statements follow from the full GeoRARE surface or its joint projection beyond target-specific intervals based on Kim and Lee (2026) or Faghihi and Saki (2026)?
  2. The main reversal and early-certification results depend on a prespecified occupancy threshold, signature, and reference masses. How stable are they to those choices? Can uncertainty over the choices be incorporated and retain simultaneous coverage?
  3. Can Theorem 3 be extended from its two-arm path construction to signed profile contrasts for a broader class of descriptors? Such a result would clarify when the unresolved-mass upper bound is rate-tight for the paper’s main targets.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.