acceptodds
Under review as a conference paper at ICLR 2027

SpecBridge: Future-Guided Spectral Alignment for Power Outage Forecasting

Abstract

Reliable power outage forecasting is critical for grid resilience, emergency response, and resource allocation, particularly as outage events can impose substantial operational and economic consequences. However, outage dynamics vary substantially across time and geographic regions, making forecasting highly challenging. Existing time-series forecasting methods primarily learn a direct mapping from historical observations to future values, but often overlook how structural relationships encoded in the past evolve toward future temporal patterns, potentially compromising forecasting accuracy as outage patterns change. We introduce SpecBridge, a future-guided spectral alignment framework for power outage forecasting. SpecBridge constructs latent graphs from numerical and textual representations of historical observations, together with a future-time graph that serves as a reference during training. It aligns the source spectral bases to this reference and adaptively aggregates them using data-driven weights that account for both source-target structural discrepancy and graph density. The aggregated historical subspace and the future-time subspace jointly supervise the forecasting representation, complementing prediction and reconstruction objectives. Empirical evaluations across 11 state-level US power outage datasets demonstrate that our SpecBridge outperforms state-of-the-art baselines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Power outage forecasting, graph transfer learning, time-series

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qXy1OpenReview ↗
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
  1. Power outage forecasting is a relevant application, and explicitly supervising temporal relationships provides an interesting alternative to aligning only latent feature values.

  2. 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
  1. 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.

  2. 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.

  3. 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.

  4. 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.

  5. 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?

  6. 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
AuthorsOpenReview ↗

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=12∥VV⊤−UfUf⊤∥F2+λ2∥VV⊤−UˉUˉ⊤∥F2.\mathcal{L}_{\mathrm{spec}} = \tfrac{1}{2}\lVert VV^{\top}-U_fU_f^{\top}\rVert_F^2 + \tfrac{\lambda}{2}\lVert VV^{\top}-\bar{U}\bar{U}^{\top}\rVert_F^2. 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.

AuthorsOpenReview ↗

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

PEMS03MSE ↓MAE ↓
SpecBridge (ours)0.06180.1664
DistDF [Wang et al., 2026]0.06800.1740
StemGNN0.09170.2036
STWave0.09240.2087
TimeBridge0.07000.1760
Fredformer0.08100.1910
iTransformer0.07200.1790
FreTS0.08500.1980
TimesNet0.09400.2010
MICN0.09600.2170
TiDE0.11700.2260
PatchTST0.09200.2100
DLinear0.10500.2200

Table 2

PEMS08MSE ↓MAE ↓
SpecBridge (ours)0.06980.1535
DistDF [Wang et al., 2026]0.07600.1770
StemGNN0.10380.2143
STWave0.10760.2261
TimeBridge0.08000.1840
Fredformer0.09100.1990
iTransformer0.08400.1870
FreTS0.09600.2050
TimesNet0.11100.2080
MICN0.16100.2740
TiDE0.12100.2330
PatchTST0.10600.2230
DLinear0.11300.2250

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).

AuthorsOpenReview ↗

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 UˉŪ and the future basis UfU_f are used as constant targets, and Lspec\mathcal{L}_spec is backpropagated only through VV. 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 XpX_p, future window XfX_f, text description Tp\mathcal{T}_p of XpX_p.

  1. Encode XpX_p by Eqs. (1) to (4) to obtain HpH_p, and compute the prediction Y^pred\hat{Y}_{pred} by Eq. (5).

  2. Apply the same procedure (Eqs. 1 to 5) to X_f to obtain HfH_f and the reconstruction Y^recon\hat{Y}_{recon}.

  3. Compute the text representation Ht=MLP(LM(Tp))H_t = MLP(LM(\mathcal{T}_p)) with the frozen CLIP text encoder and the fixed MLP (Eq. 6).

  4. For each r∈r \in {p, t, f}, build the adjacency ArA_r from HrH_r with threshold τr\tau_r (Eqs. 7 and 8) and extract its K leading eigenvectors UrU_r.

  5. For each source r∈r \in {p, t}, align UrU_r to UfU_f by orthogonal Procrustes to obtain ZrZ_r (Eq. 9), compute the discrepancy h^r\hat{h}_r and density ρ^r\hat{\rho}_r (Eq. 10) and the weight wrw_r (Eq. 11), and aggregate Uˉ=qf(∑rwrUrZr)Ū = qf(\sum_r w_r U_r Z_r) (Eq. 12). Steps 4 and 5 have no trainable parameters and carry no gradient.

  6. Rearrange the final encoder output into FxF_x, compute V=EigK(FxFxT)V = Eig_K(F_x F_xᵀ), and evaluate Lspec\mathcal{L}_{spec} with targets UfU_f and UˉŪ (Eq. 13).

  7. Update the parameters of Eqs. (1) to (5) by minimizing Ltotal=Lpred+ηreconLrecon+ηspecLspec\mathcal{L}_{total} = \mathcal{L}_{pred} + η_{recon} \mathcal{L}_{recon} + η_{spec} \mathcal{L}_{spec}. The text MLP is not updated.

Inference. Only step 1 is executed, Y^pred=fθ(Xp)Ŷ_{pred} = f_θ(X_p). 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).

StateFixed (reported) MSEFixed (reported) MAETrainable MSETrainable MAE
TX0.37030.45380.36820.4534
NY0.62530.59720.61840.5936
FL0.42910.50140.42870.5013
NC0.44650.51500.45590.5222
GA0.46680.53120.46660.5323
LA0.50840.54210.50860.5426
CA0.53310.53620.52790.5331
MI0.33250.43830.33000.4378
OH0.48530.54480.47640.5384
PA0.47610.51130.47470.5105
WA0.64880.62160.64460.6183
Average0.48380.52660.48180.5258

Table 2: Fixed MLP vs. trainable MLP, per state (with horizon = 192).

StateFixed (reported) MSEFixed (reported) MAETrainable MSETrainable MAE
TX0.45340.50130.45540.4993
NY0.77900.66970.79110.6755
FL0.47950.53610.47920.5354
NC0.49890.54940.48900.5426
GA0.51790.57030.53890.5786
LA0.56350.57480.56390.5752
CA0.63760.58900.63930.5888
MI0.43320.49860.43730.5008
OH0.55550.58400.56000.5866
PA0.59900.58090.60580.5825
WA0.81840.70480.82020.7050
Average0.57600.57810.58000.5791

(Please find the remaining answer for Q4 in Response to Reviewer qXy1 (4/8))

Reviewer_kpNdOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

SpecBridge studies power-outage time-series forecasting using future relational structure as training supervision. It constructs similarity graphs from historical numerical representations, CLIP-encoded textual summaries of the history, and future representations learned through reconstruction. Historical spectral bases are aligned to the future basis, adaptively weighted according to source-target structural discrepancy and graph density, aggregated, and orthonormalized. A spectral projection loss then guides the forecasting representation alongside prediction and reconstruction losses. Future observations are used only during training.

The paper further provides a conditional analysis of the spectral-transfer module and evaluates the method on 11 US state-level outage datasets at three forecasting horizons against 11 baselines.

Critical strengths
  1. Broad and consistent empirical improvements in the reported setting. Table 1 covers 11 states and three horizons, and the stated ranking counts are supported by the table. Several improvements are appreciable, rather than confined to tiny differences. The chronological split and training-only scaler fitting are appropriate, and Table 3 supports the usefulness of the complete spectral auxiliary objective on the nine states shown. The paper also includes seed variability, sensitivity, and computational-cost results.

  2. A coherent structural-supervision design with appropriately qualified theory. Aligning projection matrices provides a meaningful alternative to matching feature coordinates or full relation matrices. The aggregation and regularized-subspace analysis helps explain the intended role of source discrepancy and density. Remark 1 explicitly distinguishes these module-level results from end-to-end forecasting guarantees and acknowledges dependencies in learned graphs; Appendix A.3 also correctly identifies the safeguarded comparison as an oracle result.

Critical weaknesses
  1. The central graph construction and training procedure are not specified sufficiently to establish what is learned. Equation (6) describes a summary of the historical window, but the transition to NN temporally corresponding text nodes in Section 3.2 is unexplained. The mapping from numerical channel–patch embeddings to graph nodes, and the correspondence between a 720-step history and 96/192/336-step futures, also need explicit definitions. Equal node counts alone do not specify this correspondence. Further, Equation (8) uses a hard indicator: under ordinary differentiation, the text MLP receives no gradient through its only described path to the loss. This doesn't make the entire model untrainable, since the prediction features still enter the spectral loss through VV. It does require clarification of fixed components, stop-gradients, surrogate gradients, or additional training paths.

  2. The evaluation does not yet isolate the incremental contribution of adaptive spectral transfer. TimeAlign already includes future reconstruction and relational alignment, while the cited TransNet work includes Procrustes alignment, weighted aggregation, QR, and projector regularization. SpecBridge's adaptation to forecasting and its modified weighting can still be valuable, but Table 3 only removes the entire spectral objective or reconstruction objective. It does not distinguish spectral supervision from existing relational alignment, nor establish the benefits of the text source and adaptive/density-aware weighting. This matters particularly because Section 4, lines 396–401, uses recommended configurations without per-dataset tuning for baselines while searching many hyperparameters for SpecBridge. Thus, the comparison conflates method and model-selection effects. The conditional theory does not resolve this empirical attribution issue, as the authors themselves acknowledge. A compact same-backbone comparison and comparable tuning of the closest baselines would be more informative than simply expanding the benchmark count. TimeAlign §3.5; TransNet Algorithm 1.

  3. There is a smaller but concrete reporting error in the claimed percentage improvements. The percentage improvements reported in Section 4 appear to use SpecBridge's error rather than the comparison baseline's error as the denominator. For example, the TX average MSE changes from 0.4892 for TimeMixer to 0.4452 for SpecBridge. This corresponds to an 8.99%8.99\% reduction relative to TimeMixer, $ \frac{0.4892-0.4452}{0.4892}\times100%=8.99%$ rather than the reported 9.88%9.88\%. The latter value results from dividing by the SpecBridge error. The underlying Table 1 values and rankings are unaffected, so I view this as a reporting/convention issue rather than evidence against the empirical results, but the percentage statements should be defined and corrected consistently.

Questions for authors
  1. Graph/node construction and optimization. Could the authors provide one concrete tensor-shape example, for example L=720L=720, T=96T=96, and N=8N=8, including the actual channel count CC, the construction of the numerical graph nodes across channels, the textual prompt(s), the transformation of TpT_p into NN temporally aligned text-node embeddings, and the correspondence between historical and future nodes? Please also clarify which parameters receive gradients from LpredL_{\mathrm{pred}}, LreconL_{\mathrm{recon}}, and LspecL_{\mathrm{spec}}, and specifically how the text MLP is trained given the hard threshold in Eq. (8).

  2. Incremental contribution of the proposed spectral-transfer mechanism. Could the authors provide a compact matched comparison that better isolates the contribution of SpecBridge relative to the closest future-guided relational-alignment baseline? In particular, numerical-only, numerical+text with equal weighting, and the proposed adaptive weighting would help establish the contribution of the textual source and adaptive aggregation. A no-density variant would also directly test the claimed role of graph-density information if density-aware weighting remains a central contribution. A small number of carefully matched settings would be sufficient.

  3. Model selection and evaluation protocol. What validation-only selection rule and actual hyperparameter-search budget were used for SpecBridge for each dataset/horizon, and how does this compare with the tuning budget available to the strongest baselines? Please also clarify whether MSE and MAE are computed on raw customer counts, log-transformed counts, or log-plus-scaled values, and provide variability for the closest competing methods in settings where the performance differences are small. Finally, please define the convention used for the percentage improvements in Section 4 and revise those values consistently; for example, 0.4892→0.44520.4892\rightarrow0.4452 is an 8.99%8.99\% reduction relative to the TimeMixer baseline rather than 9.88%9.88\%.

14 comments
AuthorsOpenReview ↗

Response to Reviewer kpNd (1/6)

We sincerely thank the reviewer for their constructive comments, which have helped us to improve the paper's clarity and overall quality. We have uploaded an updated version of the manuscript. Below are our responses to the comments, and we hope they resolve the reviewer's concerns.

Q1: The central graph construction and training procedure. The transition to NN temporally corresponding text nodes in Section 3.2 is unexplained. The mapping from numerical channel–patch embeddings to graph nodes, and the correspondence between a 720-step history and 96/192/336-step futures, also need explicit definitions. Equal node counts alone do not specify this correspondence. Equation (8) uses a hard indicator: under ordinary differentiation, the text MLP receives no gradient through its only described path to the loss. This doesn't make the entire model untrainable, since the prediction features still enter the spectral loss through VV.

A: Thank you for your question. SpecBridge constructs latent similarity graphs over temporal instances, whose edges represent similarity between their representations rather than physical connections in the power grid. Through Eq. 7 (in the main body), the adjacency matrices are determined by the representations and the specified similarity rule; they are not independently parameterized or directly optimized.

For each historical–future window pair, we jointly minimize the prediction, reconstruction, and spectral losses. The historical patch embedding and forecasting encoder receive gradients from both the prediction and spectral losses, while the prediction head is updated by the prediction loss. The future patch embedding and reconstruction head are updated by the reconstruction loss. The language encoder and text MLP projection remain frozen. Thus, training learns forecasting representations guided by both prediction accuracy and spectral supervision, alongside a future representation learned through reconstruction.

In fact, we partition the historical window into the same NN temporal patches used by the numerical branch. For each patch ii, we construct a textual description Tp,i\mathcal{T}_{p,i} containing its local summary statistics and the shared domain context. Each description is independently encoded using the same frozen language encoder and MLP: $h_{t,i}=\text{MLP}(\text{LM}(\mathcal{T}_{p,i})) , H_t=[h_{t,1},\ldots,h_{t,N}]^\top$. Thus, text node ii and numerical node ii describe the same historical interval, and their correspondence follows directly from the shared patch partition and chronological ordering.

The encoder notation accommodates CC channels, yielding CNCN channel–patch embeddings. In our state-level experiments, each input is a single aggregate outage-count series, so C=1C=1. Reshaping the encoder output as H~r∈RC×N×D\widetilde{H}_r \in \mathbb{R}^{C \times N \times D}, graph node ii is represented by hr,i=H~r[1,i,:]h_{r,i}=\widetilde{H}_r[1,i,:]. Thus, each node represents one temporal patch, without channel pooling. We partition the historical and future windows separately into NN ordered patches. For history length L=720L=720 and forecast length TT, node ii represents Pi={(i−1)L/N+1,…,iL/N}P_i=\{(i-1)L/N+1,\ldots,iL/N\} in the history and Fi={L+(i−1)T/N+1,…,L+iT/N}F_i=\{L+(i-1)T/N+1,\ldots,L+iT/N\} in the future. For example, with N=8N=8, historical nodes each cover 90 steps, whereas future nodes cover 12, 24, or 42 steps for T=96,192,T=96,192, or 336336, respectively. The correspondence is therefore defined by relative position within each window, rather than by identical timestamps or patch durations. This indexing provides a common coordinate system for comparing within-window relational structures across different forecasting horizons. The resulting alignment acts as a structural regularizer: it encourages the historical representation to capture relationships informative about the future window, without assuming that paired patches contain identical events or outage values

We now clarify the training paths as follows. Fixed component. The text MLP is initialized randomly and kept fixed. It acts as a fixed projection of the CLIP embeddings to the model dimension, and it receives no gradient because of the hard indicator in Eq. (8). Stop-gradient. The aggregated spectral basis built from the past, text and future graphs is treated as a constant target. The spectral loss is backpropagated only through VV, that is, through the prediction features. No surrogate gradient. No surrogate or straight-through gradient is used in any reported result, and there is no other training path to the text MLP.

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.

(Please find the remaining answer for Q1 in Response to Reviewer kpNd (2/6))

AuthorsOpenReview ↗

Response to Reviewer kpNd (2/6)

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).

StateFixed (reported) MSEFixed (reported) MAETrainable MSETrainable MAE
TX0.37030.45380.36820.4534
NY0.62530.59720.61840.5936
FL0.42910.50140.42870.5013
NC0.44650.51500.45590.5222
GA0.46680.53120.46660.5323
LA0.50840.54210.50860.5426
CA0.53310.53620.52790.5331
MI0.33250.43830.33000.4378
OH0.48530.54480.47640.5384
PA0.47610.51130.47470.5105
WA0.64880.62160.64460.6183
Average0.48380.52660.48180.5258

Table 2: Fixed MLP vs. trainable MLP, per state (with horizon = 192).

StateFixed (reported) MSEFixed (reported) MAETrainable MSETrainable MAE
TX0.45340.50130.45540.4993
NY0.77900.66970.79110.6755
FL0.47950.53610.47920.5354
NC0.49890.54940.48900.5426
GA0.51790.57030.53890.5786
LA0.56350.57480.56390.5752
CA0.63760.58900.63930.5888
MI0.43320.49860.43730.5008
OH0.55550.58400.56000.5866
PA0.59900.58090.60580.5825
WA0.81840.70480.82020.7050
Average0.57600.57810.58000.5791

Table 3: Fixed MLP vs. trainable MLP, per state (with horizon = 336).

StateFixed (reported) MSEFixed (reported) MAETrainable MSETrainable MAE
TX0.51200.53250.49520.5229
NY0.83710.69980.83670.7007
FL0.49240.54140.49610.5433
NC0.51340.55730.51950.5609
GA0.51030.56770.51720.5701
LA0.58070.58690.60680.6003
CA0.70900.62550.71720.6292
MI0.51700.54100.51240.5393
OH0.55410.58650.55320.5859
PA0.64580.61460.65790.6160
WA0.96550.76850.93570.7542
Average0.62160.60200.62260.6027

Table 4: Averaged MSE and MAE.

StateFixed MLP MSE ↓Fixed MLP MAE ↓Trainable MLP MSE ↓Trainable MLP MAE ↓
TX0.44520.49590.43960.4942
NY0.74710.65560.74870.6566
FL0.46700.52630.46800.5267
NC0.48630.54060.48810.5419
GA0.49830.55640.50760.5603
LA0.55090.56790.55980.5727
CA0.62660.58360.62810.5837
MI0.42760.49260.42660.4926
OH0.53160.57180.52990.5703
PA0.57360.56890.57950.5697
WA0.81090.69830.80020.6925
Overall Avg0.56050.56890.56150.5692
AuthorsOpenReview ↗

Response to Reviewer kpNd (3/6)

Q2: The evaluation does not yet isolate the incremental contribution of adaptive spectral transfer.

A: We thank the reviewer for the important suggestion. We have added new ablation study results in the Appendix C.4 of the revised paper. To address this concern, we conducted additional experiments at a forecasting horizon of 96 across 11 states, comparing multiple configurations: prediction-only, prediction + reconstruction, relational alignment, numerical-only spectral transfer, numerical + text with equal weighting, and no-density weighting.

As shown in Tables 5-15, we observe that SpecBridge achieves the lowest MSE and MAE in all 11 states. Compared with prediction-only and prediction + reconstruction, it reduces MSE/MAE by an average of 2.5743%/2.2197% and 3.4619%/2.3425%, respectively, supporting the overall benefit of the proposed transfer mechanism beyond prediction training and future reconstruction alone. It further outperforms relational alignment, with average MSE/MAE reductions of 2.5844%/1.8512%, providing evidence for its incremental contribution beyond this alternative alignment objective. Comparisons with numerical-only spectral transfer, equal weighting, and no-density weighting yield average MSE/MAE reductions of 4.4922%/2.8770%, 4.4542%/3.0693%, and 5.0453%/3.3043%, respectively. These results support the complete adaptive spectral-transfer mechanism and the benefits of adaptive aggregation and density information under the evaluated settings.

Table 5

TXMSE ↓MAE ↓
Full SpecBridge (ours)0.37030.4538
Prediction-only0.37760.4639
Prediction + reconstruction0.39080.4703
Relational alignment0.37800.4639
Numerical-only spectral0.38720.4648
Equal weighting0.38050.4668
No-density weighting0.37540.4608

Table 6

NYMSE ↓MAE ↓
Full SpecBridge (ours)0.62530.5972
Prediction-only0.65010.6089
Prediction + reconstruction0.66650.6149
Relational alignment0.64860.6095
Numerical-only0.64290.6059
Equal weighting0.64620.6062
No-density weighting0.67330.6242

Table 7

FLMSE ↓MAE ↓
Full SpecBridge (ours)0.42910.5014
Prediction-only0.43790.5269
Prediction + reconstruction0.44260.5294
Relational alignment0.43720.5270
Numerical-only0.44170.5208
Equal weighting0.44470.5229
No-density weighting0.44290.5226

Table 8

NCMSE ↓MAE ↓
Full SpecBridge (ours)0.44650.5150
Prediction-only0.45500.5222
Prediction + reconstruction0.47000.5276
Relational alignment0.45410.5223
Numerical-only0.47460.5333
Equal weighting0.46500.5306
No-density weighting0.46780.5312

Table 9

GAMSE ↓MAE ↓
Full SpecBridge (ours)0.46680.5312
Prediction-only0.49010.5483
Prediction + reconstruction0.48370.5368
Relational alignment0.48280.5416
Numerical-only0.49550.5467
Equal weighting0.49370.5476
No-density weighting0.48940.5449

Table 10

LAMSE ↓MAE ↓
Full SpecBridge (ours)0.50840.5421
Prediction-only0.52180.5514
Prediction + reconstruction0.52160.5508
Relational alignment0.52360.5531
Numerical-only0.53390.5574
Equal weighting0.53610.5600
No-density weighting0.54390.5669

Table 11

CAMSE ↓MAE ↓
Full SpecBridge (ours)0.53310.5362
Prediction-only0.55420.5430
Prediction + reconstruction0.54910.5480
Relational alignment0.55310.5420
Numerical-only0.54880.5461
Equal weighting0.55480.5501
No-density weighting0.56980.5560

Table 12

MIMSE ↓MAE ↓
Full SpecBridge (ours)0.33250.4383
Prediction-only0.34180.4481
Prediction + reconstruction0.34590.4479
Relational alignment0.33780.4448
Numerical-only0.35060.4628
Equal weighting0.35150.4631
No-density weighting0.35000.4577

Table 13

OHMSE ↓MAE ↓
Full SpecBridge (ours)0.48530.5448
Prediction-only0.49610.5559
Prediction + reconstruction0.49540.5663
Relational alignment0.49470.5545
Numerical-only0.50510.5569
Equal weighting0.50510.5565
No-density weighting0.50890.5590

(Please find the remaining answer for Q2 in Response to Reviewer kpNd (4/6))

Reviewer_H4r6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SpecBridge, a future-guided spectral alignment framework designed for long-term power outage forecasting. To overcome the limitation that existing time-series methods focus on direct mapping while overlooking structural evolutionary relationships, SpecBridge constructs latent graphs from numerical and textual representations of historical observations, and utilizes a future-time graph as a reference during training. It employs an adaptive spectral alignment mechanism based on structural discrepancy and graph sparsity. Extensive experiments across 11 US state-level power outage datasets and 11 state-of-the-art baselines demonstrate its superior forecasting performance and computational efficiency.

Critical strengths
  1. The integration of graph spectral learning with multi-view (numerical and textual) representations to model and align temporal relational structures between the past and future is both innovative and well-grounded.

  2. The paper provides comprehensive theoretical guarantees, including error bounds for weighted source aggregation, an error-bound-oracle property, and regularized transfer analysis.

  3. The evaluation spans 11 state-level datasets, 3 prediction horizons, and 11 competitive baselines, complemented by thorough ablation studies, hyperparameter sensitivity analyses, and efficiency comparisons.

Critical weaknesses
  1. The construction of textual views relies heavily on predefined statistical summaries and textual descriptions. The sensitivity of the framework to the quality, granularity, and source of these textual prompts requires further discussion.

  2. Although efficiency metrics (Table 5) show favorable memory usage, constructing multiple latent graphs and computing spectral decompositions during training introduces additional computational overhead that could scale poorly with larger, higher-resolution networks.

Questions for authors
  1. How sensitive is SpecBridge to the choice and format of the textual descriptions? Could the authors provide an analysis or discussion on how variations in the prompt template or the pretrained language model (CLIP ViT-B/32) affect the stability of the textual graph and overall forecasting performance?

  2. Although Table 5 illustrates favorable training memory and parameter counts, how does the construction of latent graphs and spectral decomposition scale during the inference stage? Does the model require heavy graph operations when predicting new future windows?

  3. Supplement the discussion on the limitations of the proposed model framework.

5 comments
AuthorsOpenReview ↗

Response to Reviewer H4r6 (1/4)

We sincerely thank the reviewer for their thoughtful and constructive feedback. Your comments have significantly improved the clarity and rigor of our work. We have uploaded an updated version of the manuscript. Below are our responses to the comments, and we hope they resolve the reviewer's concerns.

Q1: The sensitivity of the framework to the quality, granularity, and source of these textual prompts requires further discussion. & How sensitive is SpecBridge to the choice and format of the textual descriptions? Could the authors provide an analysis or discussion on how variations in the prompt template or the pretrained language model (CLIP ViT-B/32) affect the stability of the textual graph and overall forecasting performance?

A: Thanks for raising this thought-provoking question. You can refer to Appendix C.1 in the revised PDF, and for your convenience, we have also included the corresponding content here.

We have conducted additional sensitivity experiments across 11 states and three forecasting horizons (96, 192, and 336) to examine how the information included in textual descriptions affects SpecBridge. We compared the reported five-attribute description with a reduced version containing minimum, maximum, median, and trend, and an expanded version that additionally includes standard deviation, interquartile range, and 24-hour change.

The prompt for each input window is built from the attributes in Table 1. The reported (i.e., initial submitted version) template uses the five window-level attributes listed in Section 3.2 (minimum, maximum, median, overall trend, periodicity information).

Table 1: Attributes in each prompt.

AttributeDefinitionReducedReportedExpanded
Minimumsmallest value in the window✓✓✓
Maximumlargest value in the window✓✓✓
Medianmedian of the window✓✓✓
Trendup or down, sign of the net change over the window✓✓✓
Periodicityshare of training-period steps at the same hour and month that exceed the 75th percentile of the training period, and days since the series last exceeded it✓✓
Spreadstandard deviation of the window✓
IQRinterquartile range of the window✓
24-hour changelast value minus the value 24 hours earlier✓
Number of attributes458

Additionally, we also evaluated a weather-based template that contains no outage statistics and describes the same temporal window using weather records. It describes the same window with weather records: peak wind gust, mean wind speed, total rainfall and total snowfall over the last 24 hours, and over the full window the same four plus the time since the peak gust and the temperature range.

From Tables 2, 3, and 4, we observe that, relative to the reported description, the reduced and expanded variants increase state-averaged MSE by at most 2.5% and MAE by at most 1.3% across the three horizons (here we reported the relative error change). The reported configuration achieves the lowest average MAE at every horizon and the lowest average MSE at horizons 192 and 336 among all four tested variants. At horizon 96, the Weather variant improves average MSE by only 0.06%, with essentially unchanged MAE. Individual states sometimes benefit from alternative descriptions, but expanding the attribute set does not consistently improve performance. Removing periodicity produces larger average errors at every horizon, suggesting that this attribute provides useful complementary information. These results support robustness to the tested changes in attribute selection and descriptive granularity.

Table 2: Results per state (MSE / MAE) with horizon =96

StateReported (5)Reduced (4)Expanded (8)Weather
TX0.3703 / 0.45380.3857 / 0.46540.3759 / 0.45860.3696 / 0.4542
NY0.6253 / 0.59720.6291 / 0.59950.6242 / 0.59600.6242 / 0.5966
FL0.4291 / 0.50140.4275 / 0.50040.4281 / 0.50100.4286 / 0.5010
NC0.4465 / 0.51500.4721 / 0.53080.4584 / 0.52180.4591 / 0.5232
GA0.4668 / 0.53120.4846 / 0.54270.4690 / 0.53330.4674 / 0.5318
LA0.5084 / 0.54210.5087 / 0.54350.5098 / 0.54300.5084 / 0.5423
CA0.5331 / 0.53620.5354 / 0.53840.5320 / 0.53500.5329 / 0.5361
MI0.3325 / 0.43830.3361 / 0.44190.3274 / 0.43780.3304 / 0.4371
OH0.4853 / 0.54480.4770 / 0.53900.4757 / 0.53830.4748 / 0.5376
PA0.4761 / 0.51130.4798 / 0.51400.4910 / 0.52310.4772 / 0.5135
WA0.6488 / 0.62160.6432 / 0.61910.6475 / 0.62150.6464 / 0.6211
Average0.4838 / 0.52660.4890 / 0.53040.4854 / 0.52810.4835 / 0.5268

(Please find the remaining answer for Q1 in Response to Reviewer H4r6 (2/4) and (3/4))

AuthorsOpenReview ↗

Response to Reviewer H4r6 (2/4)

Table 3: Results per state (MSE / MAE) with horizon = 192

StateReported (5)Reduced (4)Expanded (8)Weather
TX0.4534 / 0.50130.4600 / 0.50460.4701 / 0.51140.4583 / 0.5028
NY0.7790 / 0.66970.7844 / 0.67310.7934 / 0.67520.7968 / 0.6774
FL0.4795 / 0.53610.4822 / 0.53680.4850 / 0.53830.4679 / 0.5296
NC0.4989 / 0.54940.5140 / 0.55690.5012 / 0.54990.4937 / 0.5457
GA0.5179 / 0.57030.5393 / 0.58010.5390 / 0.57880.5284 / 0.5715
LA0.5635 / 0.57480.5906 / 0.59140.5780 / 0.58140.5862 / 0.5865
CA0.6376 / 0.58900.6430 / 0.59060.6418 / 0.59280.6382 / 0.5905
MI0.4332 / 0.49860.4359 / 0.49920.4283 / 0.49750.4511 / 0.5085
OH0.5555 / 0.58400.5632 / 0.58900.5563 / 0.58420.5611 / 0.5881
PA0.5990 / 0.58090.6288 / 0.59900.6146 / 0.58930.6015 / 0.5810
WA0.8184 / 0.70480.8130 / 0.70190.8211 / 0.70600.8400 / 0.7166
Average0.5760 / 0.57810.5868 / 0.58390.5844 / 0.58230.5839 / 0.5817

Table 4: Results per state (MSE / MAE) with horizon = 336

StateReported (5)Reduced (4)Expanded (8)Weather
TX0.5120 / 0.53250.5351 / 0.54340.5296 / 0.54270.5403 / 0.5466
NY0.8371 / 0.69980.8568 / 0.70840.8467 / 0.70290.8543 / 0.7050
FL0.4924 / 0.54140.4980 / 0.54290.4988 / 0.54650.4961 / 0.5461
NC0.5134 / 0.55730.5059 / 0.55440.5115 / 0.55640.5144 / 0.5576
GA0.5103 / 0.56770.5337 / 0.57970.5215 / 0.57310.5107 / 0.5681
LA0.5807 / 0.58690.6121 / 0.60180.6216 / 0.60740.6164 / 0.6043
CA0.7090 / 0.62550.7199 / 0.62930.7151 / 0.63070.7249 / 0.6363
MI0.5170 / 0.54100.5110 / 0.54320.5101 / 0.54450.5146 / 0.5428
OH0.5541 / 0.58650.5681 / 0.59500.5633 / 0.59220.5626 / 0.5916
PA0.6458 / 0.61460.6952 / 0.63310.6901 / 0.63110.6532 / 0.6204
WA0.9655 / 0.76850.9683 / 0.77330.9646 / 0.76640.9827 / 0.7763
Average0.6216 / 0.60200.6368 / 0.60950.6339 / 0.60860.6337 / 0.6086

(Please find the remaining answer for Q1 in Response to Reviewer H4r6 (3/4))

AuthorsOpenReview ↗

Response to Reviewer H4r6 (3/4)

We have replaced CLIP ViT-B/32 with two text-only encoders from a different model family, i.e., all-MiniLM-L6-v2 (a sentence encoder trained with a contrastive sentence-similarity objective), and BERT-base-uncased, trained with masked language modeling. The prompt template, data splits, seeds and all hyperparameters are unchanged. As shown in Tables 5, 6, and 7, across the 11 states, SpecBridge maintains comparable average performance when CLIP is replaced by either text-only encoder, although sensitivity increases with the forecasting horizon. Relative to CLIP, all-MiniLM-L6-v2 increases state-averaged MSE by 0.37%, 1.44%, and 3.93% at horizons 96, 192, and 336, respectively, with corresponding MAE increases of 0.25%, 0.66%, and 1.89%. BERT-base yields MSE increases of 0.95%, 1.74%, and 3.31%, and MAE increases of 0.65%, 0.78%, and 1.63%. Although alternative encoders yield lower errors for specific state–horizon combinations, CLIP achieves the lowest state-averaged MSE and MAE at all three horizons.

Table 5: Results per state (MSE / MAE) with horizon =96

StateCLIP ViT-B/32 (reported)all-MiniLM-L6-v2BERT-base
TX0.3703 / 0.45380.3719 / 0.45490.3695 / 0.4542
NY0.6253 / 0.59720.6236 / 0.59480.6189 / 0.5942
FL0.4291 / 0.50140.4349 / 0.50460.4293 / 0.5014
NC0.4465 / 0.51500.4475 / 0.51510.4539 / 0.5192
GA0.4668 / 0.53120.4651 / 0.53160.4688 / 0.5328
LA0.5084 / 0.54210.5090 / 0.54270.5256 / 0.5539
CA0.5331 / 0.53620.5306 / 0.53500.5297 / 0.5349
MI0.3325 / 0.43830.3322 / 0.43850.3443 / 0.4484
OH0.4853 / 0.54480.4755 / 0.53750.4799 / 0.5407
PA0.4761 / 0.51130.4856 / 0.51810.4756 / 0.5109
WA0.6488 / 0.62160.6658 / 0.63410.6775 / 0.6390
Average0.4838 / 0.52660.4856 / 0.52790.4884 / 0.5300

Table 6: Results per state (MSE / MAE) with horizon =192

StateCLIP ViT-B/32 (reported)all-MiniLM-L6-v2BERT-base
TX0.4534 / 0.50130.4689 / 0.50720.4761 / 0.5140
NY0.7790 / 0.66970.7721 / 0.66730.7800 / 0.6701
FL0.4795 / 0.53610.4852 / 0.53920.4674 / 0.5290
NC0.4989 / 0.54940.5042 / 0.55190.4913 / 0.5438
GA0.5179 / 0.57030.5271 / 0.57130.5236 / 0.5712
LA0.5635 / 0.57480.5922 / 0.59120.5901 / 0.5883
CA0.6376 / 0.58900.6429 / 0.59040.6392 / 0.5879
MI0.4332 / 0.49860.4416 / 0.50340.4452 / 0.5046
OH0.5555 / 0.58400.5540 / 0.58200.5633 / 0.5879
PA0.5990 / 0.58090.5988 / 0.58230.6056 / 0.5829
WA0.8184 / 0.70480.8408 / 0.71490.8639 / 0.7285
Average0.5760 / 0.57810.5843 / 0.58190.5860 / 0.5826

Table 7: Results per state (MSE / MAE) with horizon =336

StateCLIP ViT-B/32 (reported)all-MiniLM-L6-v2BERT-base
TX0.5120 / 0.53250.5143 / 0.53120.5432 / 0.5473
NY0.8371 / 0.69980.8938 / 0.72400.8570 / 0.7060
FL0.4924 / 0.54140.4900 / 0.54110.4974 / 0.5465
NC0.5134 / 0.55730.5190 / 0.56040.5157 / 0.5585
GA0.5103 / 0.56770.5257 / 0.57520.5533 / 0.5876
LA0.5807 / 0.58690.6253 / 0.61060.6350 / 0.6163
CA0.7090 / 0.62550.7446 / 0.64240.7379 / 0.6396
MI0.5170 / 0.54100.5129 / 0.54150.5225 / 0.5433
OH0.5541 / 0.58650.5661 / 0.59420.5617 / 0.5918
PA0.6458 / 0.61460.7236 / 0.64510.6740 / 0.6231
WA0.9655 / 0.76850.9904 / 0.78170.9670 / 0.7701
Average0.6216 / 0.60200.6460 / 0.61340.6422 / 0.6118

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear AC, SAC, and PC,

We thank all reviewers for their detailed reviews and insightful feedback. Their comments will help to make our work more solid and comprehensive. Below is the summary of new results and updates made based on feedback from reviewers:

Positive things

Model design and proposed methods

  • Reviewer H4r6: "The integration of graph spectral learning with multi-view (numerical and textual) representations to model and align temporal relational structures between the past and future is both innovative and well-grounded."

  • Reviewer qXy1: "Explicitly supervising temporal relationships provides an interesting alternative to aligning only latent feature values." (Reviewer qXy1 has raised the score from 2 to 3.)

  • Reviewer qXy1: "The projector-based discrepancy and Procrustes alignment appropriately handle the rotational ambiguity of spectral bases." (Reviewer qXy1 has raised the score from 2 to 3.)

  • Reviewer kpNd: "Aligning projection matrices provides a meaningful alternative to matching feature coordinates or full relation matrices."

Theoretical analysis

  • Reviewer H4r6: "The paper provides comprehensive theoretical guarantees, including error bounds for weighted source aggregation, an error-bound-oracle property, and regularized transfer analysis."

  • Reviewer kpNd: "The aggregation and regularized-subspace analysis helps explain the intended role of source discrepancy and density."

Empirical evaluation

  • Reviewer kpNd: "Broad and consistent empirical improvements in the reported setting. ... Several improvements are appreciable, rather than confined to tiny differences."

  • Reviewer H4r6: The evaluation is "complemented by thorough ablation studies, hyperparameter sensitivity analyses, and efficiency comparisons."

  • Reviewer qXy1: The paper "evaluates 11 US state outage series from 2025 over three horizons against 11 forecasting baselines, and the corresponding results are satisfactory." (Reviewer qXy1 has raised the score from 2 to 3.)

Rating score revision: We got the reply from Reviewer qXy1: "I think most of my concerns have been addressed." Reviewer qXy1 has also raised the score from 2 to 3.

Major revisions

  1. We have added component-level ablation studies comparing the full model against prediction-only, prediction + reconstruction, relational alignment, numerical-only spectral transfer, equal weighting, and no-density weighting variants. Our full model achieves the best performance across all evaluated states and horizons (Reviewer kpNd).

  2. We have extended the evaluation to traffic flow forecasting and to spectral and relational baselines, where SpecBridge remains the best method (Reviewer qXy1).

  3. We have analyzed the sensitivity to the textual view, covering prompt content, text source, and the pretrained language model, and shown that SpecBridge is robust to these choices (Reviewers H4r6 and qXy1).

  4. We have strengthened the theoretical analysis by relaxing its assumptions and proving the required concentration result, and empirically linked the theory to the learned model (Reviewer qXy1).

  5. We have clarified the motivation of the spectral loss and the trade-off between future supervision and historical regularization (Reviewer qXy1).

  6. We have clarified the training procedure, including the fixed text MLP and the text-node construction in Section 3.2, and shown that training the text MLP gives comparable results (Reviewers kpNd and qXy1).

  7. We have reported the computational cost and shown that graph operations are needed only during training (Reviewer H4r6).

  8. We have corrected the reported percentage improvements, clarified the evaluation details, and extended the discussion of limitations (Reviewers kpNd and H4r6).

Best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.