- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
FlowGMAE reconstructs unobserved edge flows on physical networks using a minimum-norm conservation solution plus a learned null-space correction. Its encoder combines incidence-aware message passing, frozen Chronos-Bolt-Tiny temporal features, and learned masking. Reconstructed flows feed a contextual deviation detector. Experiments cover two traffic districts and a simulated power network, with held-out-sensor reconstruction, anomaly detection, physics ablations, and cross-district transfer.
Critical strengths
- The decomposition is well motivated and explicitly acknowledges inconsistent noisy observations: Eq. (14) preserves the minimum achievable conservation residual, rather than claiming that arbitrary measurements admit exact balance (§3.1). Completely withholding evaluation-edge histories, including from the TSFM, is an appropriate safeguard (§3.2).
- Reconstruction gains are substantial on the reported benchmarks: D4/power nMAE improves from IGNNK's 7.53/3.61 to 4.33/1.94 (Table 1). The supplement includes useful physics-only, no-projection, and no-conservation controls (Table 5), while Table 3 reports transfer errors close to target-trained performance. Five-fold variability is reported.
Critical weaknesses
- The incremental methodological contribution is not sufficiently isolated. The cited, peer-reviewed FlowSymm already combines a minimum-norm completion with learned observation-preserving null-space corrections before a refinement step (ICLR 2026 paper, §§3.1–3.4). Section 5 understates this overlap, and no direct comparison is provided. Temporal transfer and adaptive masking could distinguish FlowGMAE, but there is no uniform-versus-learned masking ablation. Moreover, D4's physics-only result is already 4.78 versus 4.33 for the full model (Table 5), making careful attribution important.
- The anomaly-detection evidence is difficult to interpret. Section 4.2 says power anomalies are injected into held-out edges, but neither PDF specifies the intervention or whether it changes observed flows through a new physical solution. If only hidden targets change, model inputs and predictions cannot respond to the intervention. Traffic evaluation uses one matched negative window per incident and maximizes scores over [t−2,t+6], so its precision-recall results concern this constructed sample. Missing incident counts and exact matching/binning rules prevent judging the evidence adequately; continuous-monitoring performance remains untested.
Questions for authors
- For the power experiment, what changes physically, with what magnitude/duration, and which observed inputs change? Please describe whether DC power flow is re-solved and report anomaly prevalence. This determines whether Table 2 measures detection of observable physical consequences.
- What distinguishes the reconstruction mechanism from FlowSymm beyond the encoder and temporal features? A comparison under the same observations, features, and splits, or a precise explanation of incompatibility, would materially clarify the advance.
- How does the learned mask interact with the permanently withheld Esup/Eeval sets (§3.2; Appendix B)? Please specify the actual input/masked proportions and how mask-dependent projectors are updated. Does uniform masking preserve the reported gains?
3 comments
Response to reviewer L44i
Q1:
We have now clarified the power anomaly generation and evaluation detail in Section 3.2 and Appendix C.
Physical intervention. We simulated line outages for 4 hours on held-out edges . When a line is disconnected, we set flow to zero, redistribute power across the network using line outage distribution factors (LODFs). This is equivalent to re-solving DC power flow because LODFs are derived from the same linear power-flow equations, assuming the grid remains connected and power injections are fixed.
Magnitude and observability. Although outages occur on held-out edges, the power on other edges also gets redistributed. Specifically, flow on approximately 1.5% of edges changes by more than 1MW. The updated flows are provided to FlowGMAE as input to detect outage even though the failed edge is unobserved itself.
Prevalence. Across five folds, we evaluate 589 outages, with 0.28% anomaly prevalence (random AUPRC = 0.0028). This suggests Table 2 measures detection of observable physical consequences through changes in observed edges.
Q2:
Both methods start from the minimum-norm conservation solution plus corrections in null-space. However, neither FlowSymm nor FlowGMAE introduces this decomposition: it is the standard general solution of an underdetermined linear system, the basis of null-space methods in constrained optimization, and of the divergence/cycle-space split of edge flows used in graph signal processing (Lim, 2020; Jia et al. (2019)) and in observability analysis of traffic and power networks (Castillo et al. (2008)). The decoder component of our method differs in how we parameterize the null-space correction, in a nutshell:
| Property | FlowSymm | FlowGMAE (ours) |
|---|---|---|
| Architecture | GATv2 + physics solver | Graph masked autoencoder |
| Learned correction | Softmax-weighted basis coefficients | Unrestricted edge-wise residuals |
| Null-space coverage | Truncated basis (256 vectors) | Full null space, basis-independent |
| Temporal modeling | ✗ | ✓ Pretrained TSFM |
| Physics-guided adaptive masking | ✗ | ✓ |
| No known nodal injections required | ✗ | ✓ |
Decoder-only ablation (nMAE %, lower is better):
To isolate the contribution of our reconstruction mechanism, we replaced FlowGMAE's decoder with FlowSymm's null-space correction while keeping the encoder, temporal features, masking policy, and training protocol identical.
| Decoder | D4 | D7 | ACTIVSg2000 |
|---|---|---|---|
| FlowSymm-style decoder | 4.59 | 6.52 | 2.26 |
| FlowGMAE decoder (ours) | 4.33 | 6.06 | 1.94 |
| Relative improvement | 5.7% | 7.1% | 14.2% |
The table shows the part of the gain that comes from the decoder. FlowGMAE's edge-wise projected residual consistently outperforms FlowSymm across all three networks, demonstrating improvements beyond those attributable to temporal representations, graph encoding, or masking.
Conservation without known injections. Unlike FlowSymm, FlowGMAE does not require complete nodal injection measurements as inputs. FlowGMAE imposes conservation only at internal junctions and computes the right-hand side from the observed sensors alone, . Boundary nodes (unmetered ramps, loads and generators) are left free. This makes the method applicable to networks whose injections are not available, which includes both of our road networks and the power grid.
Q3:
We have now clarified our masking policy in Section 2.3 and added a uniform masking ablation in Appendix H of the revised manuscript.
Data splits. We split monitored edges (sensors) into 50% input , 40% supervision , and 10% evaluation . These sets remain fixed. Flows from are always hidden from the model and used only to compute reconstruction loss, while evaluation sensors are held out entirely.
Learnable masking. During training, we additionally hide 50% of the input sensors. We learn which sensors to mask more often based on four levels of difficulty (Q1–Q4), determined by conservation constraints. This is dynamic masking within while remains fixed.
Projector updates. We recompute the conservation matrix and projector for each new mask, caching them for reuse.
Uniform masking ablation (nMAE%):
| Masking | PeMS D4 | PeMS D7 | ACTIVSg2000 |
|---|---|---|---|
| Uniform | 4.61± 0.13 | 6.39 ± 0.11 | 2.12 ± 0.21 |
| Learnable (ours) | 4.33 ± 0.34 | 6.06 ± 0.51 | 1.94 ± 0.13 |
Our learned masking outperforms uniform masking by focusing training on edges that are harder to reconstruct.
Follow-up on Rebuttal
We hope that our additional results and revisions have addressed your concerns. As the discussion period approaches its end, we would greatly appreciate any further feedback and would be happy to clarify any remaining questions.
Thank you again for your time and consideration.
Response to reviewer L44i
We sincerely thank the reviewer for acknowledging the strengths and contributions of our work and for providing constructive feedback. We address their concerns and questions below.
We agree that the overlap with FlowSymm should be more clearly acknowledged, and will include direct comparisons in Q2.
Reconstruction mechanism. As detailed in Q2, we implemented decoder-only FlowSymm comparison under the same observations, features, and folds. This supports the benefit of our edge-wise projected residual formulation beyond the shared conservation decomposition. Also, we did a fair end-to-end comparison with FlowSymm in response to W1 of reviewer CqA7 and showed our method outperformed FlowSymm (4.33 vs. 4.62 on PeMS D4, 6.06 vs. 7.23 on PeMS D7, and 1.94 vs. 3.45 on ACTIVSg2000).
Adaptive masking. We provide a uniform-versus-learned masking ablation in Q3 along with the masking details.
Temporal representations. We also provide an additional temporal-encoder ablation in response to W3 of reviewer CqA7 comparing recurrent encoders and pretrained TSFMs.
We acknowledge that the conservation-only baseline is already strong on D4 (4.78 nMAE). Our additional experiments help isolate the contributions of the reconstruction mechanism (Q2), adaptive masking (Q3), and temporal representations (reviewer CqA7, W3). We incorporated these results and clarified the methodological distinctions in the revised manuscript.
We clarify the evaluation protocol and provide additional continuous-monitoring results below.
Power anomalies. As detailed in Q1, simulated line outages redistribute power flows across the network, including observed edges. Thus, the model inputs reflect the physical consequences of the intervention.
Traffic evaluation. The test period contains 421 (D4) and 459 (D7) reported incidents. With one matched negative per incident, the random AUPRC baseline is 0.5. We acknowledge that this balanced evaluation does not represent real-world precision, thus, we also provided continuous monitoring experiment below. We additionally report VUS-ROC and VUS-PR in our response to W3 of reviewer CqA7 to provide a complementary evaluation accounting for temporal tolerance.
Continuous monitoring experiment. To address the limitations of balanced sampling, we additionally evaluate every time step on held-out sensors. We measure the percentage of events detected within 30 minutes of onset, at a budget of one false alarm per sensor per day (one-false-alarm-per-day threshold is calibrated on validation data). For traffic, alarms more than 1 hour from reported incidents count as false alarms; for power, alarms outside simulated outage periods count as false alarms.
Unlike the matched-window evaluation, continuous monitoring does not artificially balance incident and normal samples. FlowGMAE achieves higher detection rates than the baselines on both traffic datasets at the same false-alarm budget.