acceptodds
Withdrawn from ICLR 2027

Learning Around Analytical Solutions: Risk-Calibrated Corrections for Sparse Graph Sensing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Analytical solutions can provide useful anchors for sparse graph sensing, yet learned flexibility may improve pooled error while worsening local outcomes. We ask when learning should modify or preserve an analytical anchor, whether reconstruction should correct or replace it, and whether pooled gains imply local benefits. We study two settings: a policy that makes one-sensor exchanges around a QR-selected set or retains it, and a spatiotemporal graph network that predicts a residual over Ridge reconstruction. At the smallest river budget, forced swapping improved the whole-period error point estimate but worsened 50.91% of local time-region blocks; allowing no-swap reduced harmful-block frequency by 14.95 percentage points across budgets. Under fixed QR masks, residual prediction reduced pooled unselected-field error by 18.5%–23.0% relative to direct spatiotemporal prediction, yet equal-weight local block gains remained negative on average. A three-domain comparison provided supporting evidence across river, traffic, and ocean graphs. Together, these findings show that intervention choice, correction form, and evaluation scale determine when learning improves on an analytical anchor.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Sparse Sensing, Sensor Placement, Spatiotemporal Graph Reconstruction, Learning-Augmented Algorithms

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_tgXQOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies how learning approaches in sparse graph sensing could and should be combined with analytical solutions ("anchors"). Two settings are considered: a policy that either applies a single sensor swap to a pivoted-QR sensor set or retains it, and a spatiotemporal GNN that predicts a residual over a Ridge reconstruction. The authors evaluate each method with two types of metrics: a pooled error, computed over all evaluated locations and the whole evaluation period, and a local gain, computed separately for each combination of time window and spatial region ("block") and then averaged. They find that the two can disagree. A comparison against two adapted learned systems across river, traffic, and ocean data is also reported. The authors conclude that the benefit of learning depends on whether intervention is optional, whether learning corrects or replaces the anchor, and whether it is measured by pooled or local metrics.

Critical strengths
  • Combining analytical solutions with learned corrections is a relevant direction for sparse sensing and ML in general.
Critical weaknesses
  • W1. My main concern is the presentation, which is very poor and made the paper at least for me hard to evaluate.

    • The introduction neither introduces the problem clearly nor positions the paper. It opens with a result before sparse sensing is even introduced (the 50.91% appears twice in the introduction, and ~7 times in the paper), prior work is summarized in a single sentence with four citations, and the contributions are never clearly stated.
    • Outside Sec. 2, references are almost absent and do not guide the reader. For example, pivoted QR, Ridge reconstruction, and the STGNN components ("causal temporal encoder", "relational graph blocks", "global slots") are introduced without citations or proper descriptions.
    • Many terms are undefined, e.g., "harmful block", "energy margin", "add-role support", "station-selection rule", "frozen decision", "Go", a strange versioning "v5", "v4", "v6+v4" (as in Sec. 7.3), and several acronyms.
    • The general formulation of Sec. 3 adds notation that is never used in the rest of the paper. Fig. 1 is overloaded with text in a very small font, while some of the main results (Figs. 2 and 3) are deferred to the appendix.
  • W2. I find the conclusions weak. The paper concludes that "the value of learning depends" on whether intervention is optional, whether learning corrects or replaces the anchor, and which metric is used (Sec. 8). To me, this is not a mcuh interesting finding, and the specific results are either somewhat expected or not very well supported:

    • That pooled and block-level metrics can disagree follows from their definitions, as the authors acknowledge (Sec. 7.2). That predicting a residual over a strong baseline improves over predicting from scratch is also somewhat expected (Sec. 4.3).
    • If I understood correctly, the benefit of the no-swap option comes entirely from K = 10, where keeping the QR set yields zero harmful blocks by definition, since harm is measured against that same set. At K = 25 and K = 50, the harmful-block rate increases (Sec. 6.2), and no-swap is never used in the cross-domain study (Sec. 6.5).
    • The "risk-calibrated" claim in the title is therefore fragile: whether to change the sensor set is decided by fixed, manually set thresholds, with no guarantee that a change will not make results worse (Sec. 4.2).
  • W3. As the authors acknowledge (Sec. 7.3), there is no QR+Linear or QR+Ridge baseline, so it remains unclear whether learning around the anchor helps compared with the anchor itself. Furthermore, some of their errors look implausible: Li-adapted is about four times worse than CAE-adapted on San Diego, and CAE-adapted reaches an error of 152.45 at K = 72, which is included in the mean gain.

  • W4. (Kind of related to W1) The central notion of "analytical solution" is never clearly defined. The paper only gives examples (pivoted QR and Ridge), and the operator in Sec. 3 could be any method producing a feasible solution, nothing in the definition makes it analytical, so even a neural network would qualify right? Ridge is itself fitted to historical data, and QR is applied to a data-derived matrix, so the line between analytical and learned is unclear to me.

Questions for authors
  • Q1 (related to W3). How does the complete system compare with QR+Ridge across the three domains?
  • Q2 (related to W2). In which settings, other than N264 at K = 10, does the no-swap option trigger?

Overall, I find this submission in its current state hasty and hard to follow, with conclusions that I have found it hard to understand how they go beyond "it depends". I believe it is not ready for publication at this venue.

Reviewer_H7WzOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how learning should modify, preserve, or correct analytical solutions for sparse graph sensing. It considers a learned one-sensor exchange around a pivoted-QR placement and an STGNN residual correction over Ridge reconstruction. The experiments distinguish pooled accuracy from local block-level harm and cover river, traffic, and ocean domains.

Critical strengths
  1. Treating QR placement and Ridge reconstruction as executable defaults, rather than baselines to replace entirely, is practical and clearly motivated.

  2. The paper shows that pooled error improvements can coexist with degradation in many local time-region blocks. This is an important evaluation point for sensing systems.

Critical weaknesses
  1. The “risk-calibrated” claim is not established. The placement gate combines several fixed empirical thresholds, but the paper reports neither probability calibration nor a formal or empirical risk-control guarantee. Its reduction in harmful blocks comes mainly from retaining QR at K=10K=10; at K=25K=25 and K=50K=50, the policy swaps and the harmful-block rate increases slightly. Moreover, expected-utility and expected-plus-downside ranking select the same actions, and every cross-domain setting executes a swap. The current evidence therefore supports optional intervention in one setting, but not a generally risk-calibrated correction mechanism.

  2. The cross-domain gains are not attributable to the proposed method. The Proposed, Li-adapted, and CAE-adapted systems use different selectors, decoders, capacities, training schedules, and tuning budgets, and the adaptations are not verbatim reproductions. The comparison also omits a standalone QR+Linear or QR+Ridge complete-system baseline. In addition, the reported CAE mean includes a known numerical failure at San Diego K=72K=72. Consequently, the large mean gains in Table 2 cannot isolate the benefit of the proposed placement rule, residual correction, or analytical anchoring.

Questions for authors
  1. Can the authors report calibration curves or reliability statistics for the predicted positive-transfer probability and downside score, together with the achieved harmful-block rate at each operating threshold?

  2. What is the individual contribution of each gate condition—predicted downside, positive-transfer probability, energy margin, and historical support—to swaps, abstentions, and local harm?

  3. Under a common decoder and training budget, how do QR, forced swap, and swap/no-swap placement compare across all three domains and budgets?

  4. Are the local-risk conclusions stable under alternative estimands such as absolute block-error change, worst-decile error, or CVaR, rather than block-wise relative gain?

Reviewer_yq9KOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

In sparse graph sensing, analytical methods provide reliable baselines, but learned models may improve overall accuracy while harming local performance. This paper proposes constrained learning corrections to analytical solutions. For sensor placement, QR-based selection is combined with risk gating to decide whether to swap sensors; For field reconstruction, a spatiotemporal graph neural network learns residual corrections to Ridge estimates. Experiments show fewer harmful local interventions and an 18.5%–23.0% reduction in overall error versus direct prediction. They also show that global accuracy gains do not always yield local improvements.

Critical strengths
  1. The problem is meaningful, and the analysis is detailed. The study evaluates local degradation alongside overall error across river, traffic, and ocean settings, offering useful empirical observations.
Critical weaknesses
  1. The figures are presented out of order: Figure 4 appears in the main text, while Figures 2 and 3 are placed in the appendix. Since they report results for the two core components—sensor selection and residual reconstruction—they should appear in the relevant sections.
  2. The method combines existing techniques, including QR-based sensor selection, single-sensor swapping, risk gating, and Ridge residual learning, limiting its novelty. The cross-domain experiments also omit QR+Ridge, a key analytical baseline, so the gains cannot be clearly attributed to the analytical methods or the proposed learned corrections.
Questions for authors
  1. Could QR+Ridge be added as a full-system baseline in the cross-domain experiments? How much does the proposed method improve on it?
  2. All cross-domain settings use sensor swapping. How can the effectiveness of risk gating and the no-swap mechanism be demonstrated?

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

Related papers

Discussion (0)

Sign in to comment.