acceptodds
Under review as a conference paper at ICLR 2027

FlowGMAE: Physics-Constrained Virtual Sensing for Anomaly Detection Beyond Sensor Coverage

16 views

Abstract

Real-world flow networks are only partially observed. Sensors are typically installed on only a subset of edges, leaving much of the network without direct measurements. As a result, the network state is only partially known, and anomalous events outside sensor coverage can easily go undetected. We introduce FlowGMAE, a physics-constrained graph learning framework for learning virtual sensors on partially observed networks. We make three main contributions: (1) a physics-guided graph masked autoencoder operating on the full physical network, with incidence-aware message passing and adaptive masking; (2) a decoder in which conservation directly parameterizes the reconstruction space, analytically determining the constrained component of missing flow while learning only the remaining degrees of freedom; and (3) transferable virtual sensing that combines a time-series foundation model (TSFM) with graph learning for zero-shot transfer and anomaly detection beyond sensor coverage. Experiments on physical flow networks show that FlowGMAE substantially improves flow reconstruction over diverse baselines, with particularly strong gains as the distance from observed sensors increases. Moreover, anomalies on unobserved edges remain detectable from the reconstructed flows, demonstrating the potential of virtual sensing to extend monitoring beyond the physical sensor network.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph neural networks, Physics-informed machine learning, Spatiotemporal learning, Missing data imputation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_L44iOpenReview ↗
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
  1. 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).
  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
  1. 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.
  2. 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
  1. 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.
  2. 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.
  3. 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
AuthorsOpenReview ↗

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.

W1:

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.

W2:

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.

MethodPeMS D4PeMS D7ACTIVSg2000
FlowGMAE55%59%78%
GNN + physics29%27%59%
GraphMAE226%25%51%

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.

AuthorsOpenReview ↗

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 (Eeval)(E_{\mathrm{eval}}). 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:

PropertyFlowSymmFlowGMAE (ours)
ArchitectureGATv2 + physics solverGraph masked autoencoder
Learned correctionSoftmax-weighted basis coefficientsUnrestricted edge-wise residuals
Null-space coverageTruncated 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.

DecoderD4D7ACTIVSg2000
FlowSymm-style decoder4.596.522.26
FlowGMAE decoder (ours)4.336.061.94
Relative improvement5.7%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, ct=−BintMqtobs\mathbf{c}_t = -\mathbf{B}{\mathrm{int}}\mathbf{M}\mathbf{q}_t^{\mathrm{obs}}. 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 EinputE_{\mathrm{input}}, 40% supervision EsupE_{\mathrm{sup}}, and 10% evaluation EevalE_{\mathrm{eval}}. These sets remain fixed. Flows from EsupE_{\mathrm{sup}} 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 EinputE_{\mathrm{input}} while EsupE_{\mathrm{sup}} remains fixed.

Projector updates. We recompute the conservation matrix AA and projector PP for each new mask, caching them for reuse.

Uniform masking ablation (nMAE%):

MaskingPeMS D4PeMS D7ACTIVSg2000
Uniform4.61± 0.136.39 ± 0.112.12 ± 0.21
Learnable (ours)4.33 ± 0.346.06 ± 0.511.94 ± 0.13

Our learned masking outperforms uniform masking by focusing training on edges that are harder to reconstruct.

AuthorsOpenReview ↗

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.

Reviewer_8BZeOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FlowGMAE, a physics-constrained graph masked autoencoder for reconstructing flows on partially observed networks. It combines conservation-based reconstruction, graph learning, and pretrained time-series representations to estimate flows on unobserved edges. Experiments on traffic and power networks demonstrate improved flow reconstruction, anomaly detection, and zero-shot transfer performance.

Critical strengths
  1. Combining physical conservation constraints with graph learning provides a reasonable approach to reconstructing flows beyond sensor coverage. The null-space projection effectively incorporates physical constraints into the reconstruction process.

  2. The experiments cover multiple traffic and power networks, with comparisons against diverse baselines. The improvements in flow reconstruction, particularly under sparse sensing conditions, and the zero-shot transfer results demonstrate the practical potential of the framework.

Critical weaknesses
  1. The framework assumes instantaneous flow conservation at internal junctions. However, real traffic measurements are aggregated over time, and congestion or incidents may introduce vehicle accumulation and propagation delays. The validity of this assumption under such conditions is not sufficiently examined.
Questions for authors

How does the proposed conservation constraint behave under traffic congestion or incidents where vehicle accumulation and propagation delays may occur?

2 comments
AuthorsOpenReview ↗

Response to reviewer 8BZe

We sincerely thank the reviewer for highlighting the strengths of our approach. We appreciate their thoughtful question regarding the conservation assumption, which we address below:

W1 & Q1:

1. Conservation under inconsistent observation. First, our traffic graphs include only motorways and ramps, where conservation is applied at junctions rather than along road segments. Importantly, FlowGMAE does not force exact conservation when measured inflow and outflow don’t match. The least-squares solution minimizes the conservation residual, while the learned null-space correction doesn’t increase it:

∥Ax^t−ct∥2=∥A(xtp+Pr~t)−ct∥2\|\mathbf{A}\hat{\mathbf{x}}_t-\mathbf{c}_t\|_2 = \|\mathbf{A}(\mathbf{x}_t^p+\mathbf{P}\tilde{\mathbf{r}}_t)-\mathbf{c}_t\|_2

=∥Axtp−ct∥2= \|\mathbf{A}\mathbf{x}_t^p-\mathbf{c}_t\|_2

=min⁡x∥Ax−ct∥2= \min_{\mathbf{x}}\|\mathbf{A}\mathbf{x}-\mathbf{c}_t\|_2

Thus, measurement inconsistencies are not forced to zero.

2. Conservation residuals during congestion. We examined junctions where all connected roads have sensors. We measured the inflow–outflow imbalance normalized by junction inflow capacity. We define congestion as occupancy>15%. We set the tolerance to the 95th percentile of training residuals (ϵ\epsilon=0.134 for D4 and ϵ\epsilon=0.215 for D7).

MetricD4 NormalD4 CongestedD7 NormalD7 Congested
Median normalized residual (%)9.88.99.110.2
Within tolerance ϵ\epsilon (%)95.388.995.592.5

The median residual remains around 9–10% in both conditions. Even during congestion, 89–93% of measurements remain within the tolerance. Therefore, conservation is a reasonable approximation.

3. Reconstruction accuracy during congestion. We additionally evaluated imputation performance on held-out sensors, for both normal and congested conditions (nMAE %, lower is better).

MethodD4 NormalD4 CongestedD7 NormalD7 Congested
MLP10.5018.509.4512.97
GraphMAE28.3514.598.1111.01
IGNNK7.3112.137.069.51
KITS7.3313.786.9310.70
GNN + physics4.768.835.819.31
FlowGMAE4.148.135.739.07

All methods perform worse during congestion, however, FlowGMAE remains competitive and outperforms all baselines.

4. Effectiveness during real incidents. Figure 5 illustrates a real CHP-reported incident where FlowGMAE successfully reconstructs the flow drop at an unobserved sensor. Table 2 further reports incident-detection AUROC of 0.86 on D4 and 0.78 on D7, supporting the usefulness of reconstructed flows for detecting real disruptions.

Conclusion: we believe these results show that even during congestion, conservation remains a reasonable approximation. Although FlowGMAE does not explicitly model vehicle accumulation or propagation delays, its least-squares formulation accommodates nonzero conservation residuals, and the empirical results demonstrate the benefit of conservation-guided reconstruction under congestion. We will include these analyses and clarify this limitation in the revised manuscript.

AuthorsOpenReview ↗

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.

Reviewer_CqA7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

This paper introduces FlowGMAE, which aims to mitigate the problem of partial observation in graph-based time-series anomaly detection. The whole pipeline is built on an existing foundation model for time-series graphs. Their method uses graph imputation to synthesize values for missing sensors and applies an anomaly scoring head to perform anomaly detection.

Critical strengths
  1. This work has good social benefits, especially for some CPS domains, such as water networks, where sensors are expensive and observations are often partial. 
  2. The motivation figure and evaluation figure look good.
Critical weaknesses
  1. The whole idea is built on an existing time-series graph foundation model, and graph imputation is not a new concept. The novelty of the proposed method is marginal. The performance improvement of the method is also marginal compared to common GNNs with physical constraints. 

  2. The details of the imputation process are not mentioned. It is unclear what kind of data it reconstructs, whether normal or mixed data, and how these signals help with leakage detection. 

  3. The evaluation setting is unclear. The paper claims that this is a zero-shot setting. However, what is the training dataset, and how is the method tested on the testing dataset? The paper also does not mention which foundation model they chose or what data it was trained on. The evaluation metrics are outdated. Some recent useful metrics, such as VUS-PR, would be helpful for evaluating the model's performance. 

  4. The problem formulation is hard to read. The task formulation should be in the Method section. The notation system is chaotic. Some variables are undefined, such as (a_t^{obs}) in Equation (5), and some variables are defined twice around Line 121 and Equation (5).

Questions for authors

The evaluation setting is unclear. The paper claims that this is a zero-shot setting. However, what is the training dataset, and how is the method tested on the testing dataset? The paper also does not mention which foundation model they chose or what data it was trained on. The evaluation metrics are outdated. Some recent useful metrics, such as VUS-PR, would be helpful for evaluating the model's performance.

4 comments
AuthorsOpenReview ↗

Response to reviewer CqA7

We sincerely thank the reviewer for the insightful feedback. We clarify all the points and address each point below.

W1:

We wanted to clarify FlowGMAE is not built on a time-series graph foundation model. Its graph model is trained from scratch. We used publicly-available Chronos-Bolt-Tiny which is pretrained on diverse timeseries datasets as a temporal feature extractor (Appendix C). Without the temporal module FlowGMAE still beats all baselines (nMAE 4.34 / 6.63 / 2.63 vs. 7.53 / 7.30 / 3.61 for the best baseline).

The novelty is how physics is used. Our contribution is:

(i) The reconstruction is guaranteed to conserve flow by construction, rather than encouraged to by a loss term.

(ii) The network learns only the part of the flow that conservation leaves undetermined.

(iii) It needs no knowledge of nodal injections, unlike recent prior work FlowSymm.

The average gain over best physics based baselines is 17.8% for imputation, we have also added a FlowSymm as new baseline in response to Q2 of reviewer L44i.

FlowGMAE's improvement over physics baselines.

MethodPeMS D4PeMS D7ACTIVSg2000
GNN + physics7.557.544.88
FlowSymm4.627.233.45
FlowGMAE4.336.061.94
Improvement over GNN + physics42.6%19.6%60.2%
Improvement over FlowSymm6.3%16.2%43.8%

W2:

We clarify imputation and anomaly-detection process as follows:

1. Imputation: FlowGMAE reconstructs timeseries edge flows. It artificially masks a subset of training data in EobsE_{\mathrm{obs}} and minimizes the objective on held-out EsupE_{\mathrm{sup}} (Eq. 15). This mimics missing sensors and enables the model to reconstruct unobserved flows without using any anomaly labels, data might contain both normal and anomalous flows.

2. Anomaly-detection: Our reconstruction is conditioned on observed flows:

$ \hat{\mathbf{x}}_t=f\theta(\mathbf{q}_t^{\mathrm{obs}},\mathbf{G},\mathbf{M}), $

where qtobs\mathbf{q}_t^{\mathrm{obs}} denotes observed sensor measurements, G\mathbf{G} the network graph, and M\mathbf{M} the observation mask. Thus, if an unobserved edge has an anomaly, its effects can propagate to neighboring observed flows and be reflected in the reconstruction. We conducted additional experiments detailed in our response to Reviewer L44i (W2 and Q1).

3. Evaluation. Figure 5 illustrates an incident-related flow drop, while Table 2 reports downstream anomaly-detection performance.

We emphasize that imputation and anomaly detection are separate tasks. Our experiments focus on traffic incidents and power-network anomalies, however our method could be extended to other domains, such as leakage detection in water networks.

AuthorsOpenReview ↗

Response to reviewer CqA7

W3:

Main setting is not zero-shot. Tables 1 and 2 train and test on the same network, with a chronological 70/10/20 split (Appendix C) and five folds of held-out sensors that never enter training (§2.2; Appendix C).

- Zero-shot (§3.3, Table 3): We provide a zero-shot transfer experiment in Table 3 to demonstrate our model's generalizability across networks. Physics-constrained learning and the use of a pretrained TSFM as the temporal module help the model generalize to unseen domains without retraining. Only the conservation-related terms are recomputed for the new network.

- Foundation model: As mentioned in Appendix B, we used frozen Chronos-Bolt-Tiny (Ansari et al., 2024; Appendix C, line 776), however, we revised the manuscript by clearly stating this fact in the main paper. We only used TSFM as a domain invariant temporal feature extractor. Also, we now present an ablation on the foundation model used, added to appendix F in revised manuscript:

Ablation: Performance under various time-series foundation models.

Temporal encoderPeMS D4 nMAE (%) ↓PeMS D4 R² ↑PeMS D7 nMAE (%) ↓PeMS D7 R² ↑
RNN4.51 ± 0.390.820 ± 0.0576.48 ± 0.620.631 ± 0.072
GRU4.46 ± 0.360.825 ± 0.0556.41 ± 0.590.637 ± 0.069
LSTM4.49 ± 0.380.822 ± 0.0566.44 ± 0.600.634 ± 0.070
Chronos-Bolt-Tiny (frozen)4.33 ± 0.340.839 ± 0.0536.06 ± 0.510.659 ± 0.066
Chronos-Bolt-Small (frozen)4.36 ± 0.350.835 ± 0.0546.11 ± 0.530.655 ± 0.067
Chronos-T5-Small (frozen)4.40 ± 0.370.831 ± 0.0556.18 ± 0.550.649 ± 0.068
Chronos-2 (frozen)4.38 ± 0.360.833 ± 0.0546.14 ± 0.540.653 ± 0.067

We chose Chronos-Bolt-Tiny as the best performer.

-Metric: We thank the reviewer for this suggestion. We initially reported AUROC and AUPRC to remain consistent with established evaluation practices in time-series anomaly detection, including recent work such as CAROTS (2025), which reports both metrics, and CANDI (AAAI 2026), which reports AUROC. We agree that these metrics alone do not fully capture detection quality. We have therefore extended our evaluation to include VUS-ROC and VUS-PR.

Additional results. VUS-ROC and VUS-PR across three datasets.

MethodD4 VUS-ROCD4 VUS-PRD7 VUS-ROCD7 VUS-PRACTIVSg2000 VUS-ROCACTIVSg2000 VUS-PR
MLP0.5510.5080.5360.5040.5210.003
GraphMAE20.6720.5930.6450.5610.6310.011
IGNNK0.7410.6610.7130.6280.7020.019
KITS0.7240.6430.7010.6150.6840.016
GNN + physics0.7590.6740.7310.6420.7390.024
FlowGMAE0.8310.7520.8060.7180.8470.041

ACTIVSg2000 anomaly prevalence: 0.28% (random AUPRC = 0.0028).

W4:

We have revised the manuscript by moving the task formulation into Section 3 (Methodology), explicitly defining all variable (qtobs,Xtobsq_t^{\mathrm{obs}}, X_t^{\mathrm{obs}}) before use, resolving all duplicate/conflicting variables (e.g., WW vs. KK, btb_t vs. ff), and we also added a full notation table in Appendix A.

Q1:

Please refer to W3.

AuthorsOpenReview ↗

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.

Thank you for the detailed response and for addressing my concerns. I appreciate the clarifications and additional results, and I will raise my score accordingly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author response to all reviewers

We sincerely thank all the reviewers for thoughtful feedback and recognizing the motivation and practical value of our work. The main concerns were (i) novelty and comparison with recent prior work (CqA7, L44i), (ii) conservation assumptions under congestion (8BZe), and (iii) anomaly-detection evaluation and experimental details (CqA7, L44i).

To address these concerns, we provided these additional details:

Novelty and baseline comparison: Clarified that FlowGMAE's graph model is trained from scratch and uses a frozen TSFM only as a temporal feature extractor. Added end-to-end and decoder-only comparisons with FlowSymm, demonstrating consistent improvements across all three datasets (CqA7, L44i).

Ablation studies: Added uniform-versus-learned masking ablation and TSFM encoder comparisons to isolate the contributions of our components (CqA7, L44i).

Conservation under congestion: Measured and analyzed conservation residuals and reconstruction performance under normal-versus-congested conditions, demonstrating the effectiveness of our method despite nonzero conservation residuals (8BZe).

Expanded anomaly evaluation: Added VUS-ROC, VUS-PR metrics, and continuous-monitoring experiments to strengthen the anomaly-detection evidence (CqA7, L44i).

Experimental details: Clarified the zero-shot evaluation setting, the power-outage intervention and flow redistribution, and the masking and data-split protocol (CqA7, L44i).

We also revised the manuscript and improved problem formulation, notation, methodological distinction, and experimental evidence.

We hope these additional results in our individual responses and the revision address the reviewers' concerns. As the discussion period approaches its end, we would greatly appreciate any further feedback and would be happy to clarify any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.