- 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.