- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes SpecBridge, which forecasts power outages using patch-based attention and frozen CLIP representations of historical summaries. SpecBridge adopts a future-guided spectral alignment framework to obtain reliable representations for future prediction. This idea is simple and motivative. The paper gives conditional spectral-transfer bounds and evaluates 11 US state outage series from 2025 over three horizons against 11 forecasting baselines, and the corresponding results are satisfactory.
Critical strengths
-
Power outage forecasting is a relevant application, and explicitly supervising temporal relationships provides an interesting alternative to aligning only latent feature values.
-
The projector-based discrepancy and Procrustes alignment appropriately handle the rotational ambiguity of spectral bases. The paper clearly limits its theoretical guarantees to the spectral module.
Critical weaknesses
-
First, the alignment loss in Eq.(13) is somewhat confusing, and it is given without motivation. The third term is indeed the alignment term, but the meaning of the second term is missing. Moreover, why must the proposed SpecBridge align the eigenvectors w.r.t. past and future relational matrices? The relation of representations for past and future data may temporally evolve across time, even though the nodes are kept identical, so the motivation to directly align structure of representations risks learning plasticity of new power outage patterns. Additionally, the power outage is always related to human events, and these events can make relations vary with randomness that is hard to predict.
-
All the assumptions adopted in theoretical analysis require more stringent verifications. Can they be easily satisfied practically on the datasets considered? The spectral theory assumes rather than establishes the difficult statistical premise for the implemented graphs. Assumption 1 uses independent Bernoulli edges and independent graph layers, whereas the actual edges arise from jointly learned, correlated embeddings of the same series. More importantly, Assumption 3 Eq. (21) directly supplies uniform aggregate-subspace concentration and comparability of empirical and population weighting scores. The proofs use these premises, but the claimed first-order concentration justification is not derived. Please verify if all the assumptions can be satisfied in practice.
-
Can the proposed SpecBridge method be applicable to more scenarios, such as traffic flow forecasting (e.g., METR-LA, PEMS04, PEMS07, and PEMS08 datasets)? Moreover, additional baseline methods about time series forecasting that use spectral decomposition or relational analysis of nodes should be considered. This will make the method more convincing for more general time series forecasting problems.
-
In Eq.(6), a trainable MLP is adopted, but H_t is then adopted to get A_t and U_t, so can the parameters of this MLP still be updated in the processing process? If the adjacency and eigenspaces are detached, this MLP appears to have no described training signal. It is better if the authors can provide a complete and detailed algorithm description.
-
The proposed method additionally uses text information to guide future prediction. Do the adopted baseline methods all use the same data for forecasting, or if only limited text information is provided, can this method still be useful and applicable? Once limited text information is provided, the graph would be degenerate; how can SpecBridge solve this problem?
-
The regularization theorem is not linked to an implementable selection rule or the learned network. Theorem 2 chooses lambda using unknown target concentration and population heterogeneity, and its safeguarded endpoint comparison is explicitly an oracle. It also analyzes the exact global minimizer over arbitrary K-dimensional subspaces, while V in Eq. (13) is constrained to come from forecasting features and is trained jointly with other losses. It would be better if the authors can verify the gap between theoretical results and the practical implementation.
Questions for authors
see weaknesses
10 comments
Response to Reviewer qXy1 (2/8)
Q3: Can the proposed SpecBridge method be applicable to more scenarios, such as traffic flow forecasting (e.g., METR-LA, PEMS04, PEMS07, and PEMS08 datasets)? Moreover, additional baseline methods about time series forecasting that use spectral decomposition or relational analysis of nodes should be considered. This will make the method more convincing for more general time series forecasting problems.
A: We thank the reviewer for this suggestion. We have conducted additional experiments on PEMS03 and PEMS08, using a forecasting horizon of 12 steps. You can refer to Appendix C.5 in the revised PDF, and for your convenience, we have also included the corresponding content here.
We also expand the comparison to include StemGNN [Cao et al., 2020] and STWave [Fang et al., 2023], which directly address the reviewer’s request for spectral and relational baselines. StemGNN jointly models temporal dependencies and cross-series relationships through discrete and graph Fourier transforms, while STWave combines wavelet decomposition with spectral graph attention. Together with FreTS and Fredformer, these methods provide a more targeted comparison with existing frequency-domain approaches.
As shown in Tables 1 and 2, SpecBridge achieves the lowest MSE and MAE on both datasets. Compared with the runner-up in each case, SpecBridge reduces MSE and MAE by 9.12% and 4.37% on PEMS03, and by 8.16% and 13.28% on PEMS08, respectively. Averaging the relative reductions equally across the two datasets yields improvements of 8.64% in MSE and 8.82% in MAE. These results support the applicability of SpecBridge to traffic flow forecasting and demonstrate its competitive performance against both general forecasting methods and methods explicitly modeling spectral structure and node relationships.
Table 1
| PEMS03 | MSE ↓ | MAE ↓ |
|---|---|---|
| SpecBridge (ours) | 0.0618 | 0.1664 |
| DistDF [Wang et al., 2026] | 0.0680 | 0.1740 |
| StemGNN | 0.0917 | 0.2036 |
| STWave | 0.0924 | 0.2087 |
| TimeBridge | 0.0700 | 0.1760 |
| Fredformer | 0.0810 | 0.1910 |
| iTransformer | 0.0720 | 0.1790 |
| FreTS | 0.0850 | 0.1980 |
| TimesNet | 0.0940 | 0.2010 |
| MICN | 0.0960 | 0.2170 |
| TiDE | 0.1170 | 0.2260 |
| PatchTST | 0.0920 | 0.2100 |
| DLinear | 0.1050 | 0.2200 |
Table 2
| PEMS08 | MSE ↓ | MAE ↓ |
|---|---|---|
| SpecBridge (ours) | 0.0698 | 0.1535 |
| DistDF [Wang et al., 2026] | 0.0760 | 0.1770 |
| StemGNN | 0.1038 | 0.2143 |
| STWave | 0.1076 | 0.2261 |
| TimeBridge | 0.0800 | 0.1840 |
| Fredformer | 0.0910 | 0.1990 |
| iTransformer | 0.0840 | 0.1870 |
| FreTS | 0.0960 | 0.2050 |
| TimesNet | 0.1110 | 0.2080 |
| MICN | 0.1610 | 0.2740 |
| TiDE | 0.1210 | 0.2330 |
| PatchTST | 0.1060 | 0.2230 |
| DLinear | 0.1130 | 0.2250 |
Cao, D., Wang, Y., Duan, J., Zhang, C., Zhu, X., Huang, C., Tong, Y., Xu, B., Bai, J., Tong, J. and Zhang, Q., 2020. Spectral temporal graph neural network for multivariate time-series forecasting. Advances in neural information processing systems, 33, pp.17766-17778.
Fang, Y., Qin, Y., Luo, H., Zhao, F., Xu, B., Zeng, L. and Wang, C., 2023, April. When spatio-temporal meet wavelets: Disentangled traffic forecasting via efficient spectral graph attention networks. In 2023 IEEE 39th international conference on data engineering (ICDE) (pp. 517-529). IEEE.
Wang, H., Pan, L., Lu, Y., Chu, Z., Li, X., He, S., Chan, Z., Wen, Q., Li, H. and Lin, Z., 2026, April. Distdf: Time-series forecasting needs joint-distribution wasserstein alignment. In International Conference on Learning Representations (Vol. 2026, pp. 131512-131535).
Response to Reviewer qXy1 (3/8)
Q4: In Eq.(6), a trainable MLP is adopted, but H_t is then adopted to get A_t and U_t, so can the parameters of this MLP still be updated in the processing process? If the adjacency and eigenspaces are detached, this MLP appears to have no described training signal. It is better if the authors can provide a complete and detailed algorithm description.
A: The parameters of the MLP in Eq. (6) are not updated. Because of the indicator in Eq. (8), no gradient reaches this MLP, so it is initialized randomly and kept fixed, and it acts as a fixed projection of the text embedding to the model dimension. The aggregated basis and the future basis are used as constant targets, and is backpropagated only through . We now state this explicitly in Section 3.2 and have added the following algorithm to the revision.
Algorithm 1. Training SpecBridge (one mini-batch).
Input: past window , future window , text description of .
-
Encode by Eqs. (1) to (4) to obtain , and compute the prediction by Eq. (5).
-
Apply the same procedure (Eqs. 1 to 5) to X_f to obtain and the reconstruction .
-
Compute the text representation with the frozen CLIP text encoder and the fixed MLP (Eq. 6).
-
For each {p, t, f}, build the adjacency from with threshold (Eqs. 7 and 8) and extract its K leading eigenvectors .
-
For each source {p, t}, align to by orthogonal Procrustes to obtain (Eq. 9), compute the discrepancy and density (Eq. 10) and the weight (Eq. 11), and aggregate (Eq. 12). Steps 4 and 5 have no trainable parameters and carry no gradient.
-
Rearrange the final encoder output into , compute , and evaluate with targets and (Eq. 13).
-
Update the parameters of Eqs. (1) to (5) by minimizing . The text MLP is not updated.
Inference. Only step 1 is executed, . Steps 2 to 6 require the future window and are used only in training.
To check whether training the text MLP would change the results, we ran a variant in which the indicator of the text graph is replaced by a straight-through estimator, so that the text MLP is trained through the spectral loss. Everything else is unchanged.
Tables 1, 2, and 3 compare the fixed and trainable variants across all 11 states and three forecast horizons (96, 192, 336). The state-averaged MSE and MAE differ by at most approximately 0.70% and 0.18%, respectively, with no consistent advantage for the trainable variant. These results support retaining the fixed text MLP in the reported configuration.
In addition, averaging over the three forecast horizons for each state (see Table 4), the fixed text MLP achieves lower MSE in 7 of 11 states and lower MAE in 7 states, with one MAE tie. Across all states and horizons, making the text MLP trainable increases the overall mean MSE and MAE by 0.18% and 0.05%, respectively. We have added these new results in the Appendix C.3 of the revised paper.
Table 1: Fixed MLP vs. trainable MLP, per state (with horizon = 96).
| State | Fixed (reported) MSE | Fixed (reported) MAE | Trainable MSE | Trainable MAE |
|---|---|---|---|---|
| TX | 0.3703 | 0.4538 | 0.3682 | 0.4534 |
| NY | 0.6253 | 0.5972 | 0.6184 | 0.5936 |
| FL | 0.4291 | 0.5014 | 0.4287 | 0.5013 |
| NC | 0.4465 | 0.5150 | 0.4559 | 0.5222 |
| GA | 0.4668 | 0.5312 | 0.4666 | 0.5323 |
| LA | 0.5084 | 0.5421 | 0.5086 | 0.5426 |
| CA | 0.5331 | 0.5362 | 0.5279 | 0.5331 |
| MI | 0.3325 | 0.4383 | 0.3300 | 0.4378 |
| OH | 0.4853 | 0.5448 | 0.4764 | 0.5384 |
| PA | 0.4761 | 0.5113 | 0.4747 | 0.5105 |
| WA | 0.6488 | 0.6216 | 0.6446 | 0.6183 |
| Average | 0.4838 | 0.5266 | 0.4818 | 0.5258 |
Table 2: Fixed MLP vs. trainable MLP, per state (with horizon = 192).
| State | Fixed (reported) MSE | Fixed (reported) MAE | Trainable MSE | Trainable MAE |
|---|---|---|---|---|
| TX | 0.4534 | 0.5013 | 0.4554 | 0.4993 |
| NY | 0.7790 | 0.6697 | 0.7911 | 0.6755 |
| FL | 0.4795 | 0.5361 | 0.4792 | 0.5354 |
| NC | 0.4989 | 0.5494 | 0.4890 | 0.5426 |
| GA | 0.5179 | 0.5703 | 0.5389 | 0.5786 |
| LA | 0.5635 | 0.5748 | 0.5639 | 0.5752 |
| CA | 0.6376 | 0.5890 | 0.6393 | 0.5888 |
| MI | 0.4332 | 0.4986 | 0.4373 | 0.5008 |
| OH | 0.5555 | 0.5840 | 0.5600 | 0.5866 |
| PA | 0.5990 | 0.5809 | 0.6058 | 0.5825 |
| WA | 0.8184 | 0.7048 | 0.8202 | 0.7050 |
| Average | 0.5760 | 0.5781 | 0.5800 | 0.5791 |
(Please find the remaining answer for Q4 in Response to Reviewer qXy1 (4/8))
Response to Reviewer qXy1 (1/8)
We thank you for your thorough review and appreciate both the positive feedback and the constructive comments. Your comments have significantly improved the clarity and rigor of our work. We have uploaded an updated version of the manuscript. Please find our responses to your points below.
Q1: First, the alignment loss in Eq.(13) is somewhat confusing, and it is given without motivation.
A: We agree that the forecasting model should retain the flexibility to learn new outage patterns as relationships among temporal patches evolve. As the reviewer points out, the third term aligns the forecasting subspace V with the aggregated historical subspace Ū. The first two terms, the constant K and the negative trace term, together have the same form and align V with the observed future subspace U_f. In the revised manuscript, we have added the following equivalent form to Eq. (13) and clarified the role of each term and the trade-off controlled by λ in the paragraph immediately following the equation.
Lspec=21∥VV⊤−UfUf⊤∥F2+2λ∥VV⊤−UˉUˉ⊤∥F2. The first term uses the observed future graph to provide structural supervision during training. The second term encourages the forecasting features to retain historical relationships that may remain informative in the forecast window. Both terms compare subspaces through their projection matrices. The parameter λ controls the balance between future structural supervision and historical regularization. A smaller λ gives the forecasting subspace more freedom to depart from historical structure. This flexibility lets the model learn emerging relational patterns when the historical inputs contain predictive information.
Q2: All the assumptions adopted in theoretical analysis require more stringent verifications. Can they be easily satisfied practically on the datasets considered? The spectral theory assumes rather than establishes the difficult statistical premise for the implemented graphs. Assumption 1 uses independent Bernoulli edges and independent graph layers, whereas the actual edges arise from jointly learned, correlated embeddings of the same series. More importantly, Assumption 3 Eq. (21) directly supplies uniform aggregate-subspace concentration and comparability of empirical and population weighting scores. The proofs use these premises, but the claimed first-order concentration justification is not derived. Please verify if all the assumptions can be satisfied in practice.
A: For Assumption 1, we agree that assuming independent Bernoulli edges and independent graph layers is restrictive for graphs constructed from jointly learned embeddings. We have therefore removed both independence requirements. The revised analysis allows dependence among edges and across graph layers, provided that the bounds in Eq. (20) hold jointly on a high-probability event. For Eq. (21), we agree that both claims require justification. We have therefore moved Eq. (21) out of Assumption 3 and proved both claims in a new lemma. For a fixed number of source graphs, the single-graph bounds in Eq. (20), together with the spectral regularity conditions, yield the aggregate-subspace perturbation bound uniformly over the weight simplex and establish the comparability of the empirical and population weighting scores. The revised assumption and Lemma 1 are in Appendix A.1, and the proof is in Appendix A.2.
After these revisions, Eq. (20) is the main probabilistic assumption. For independent edges, its Bernstein-scale bounds follow from standard concentration and eigenspace perturbation under suitable degree and eigengap conditions [1]. Under the same conditions, Corollary 3.5 of [2] gives the same concentration rates for dependent edges satisfying two-sided ℓ∞-independence with a uniformly bounded parameter.
[1] Tropp, J. A. (2012). User-friendly tail bounds for sums of random matrices. Foundations of Computational Mathematics, 12, 389–434.
[2] Adamczak, R., and Kavvadias, I. (2025). Matrix concentration inequalities for dependent binary random variables. arXiv:2504.08138.