acceptodds
Under review as a conference paper at ICLR 2027

TVTS: Learning Time-Varying Dynamics and Directed Cross-Variable Dependencies for Multivariate Time Series Forecasting

Abstract

Multivariate Time Series Forecasting (MTSF) requires jointly modeling both temporally heterogeneous patterns and cross-variable dependencies. However, many recent designs rely on time-invariant temporal filters and symmetric similarity for variable interactions, which limits the faithful representation of position-dependent temporal dynamics and directional information propagation. To address these challenges, we propose the Time-Varying Directed Geometry framework for Multivariate Time Series Forecasting (TVTS). For temporal modeling, we introduce a Time-Varying Convolutional Filter that learns distinct spectral responses across relative temporal positions through a structured basis expansion, thereby relaxing the shift-invariance assumption of conventional filters. For cross-variable modeling, we propose Directed Geometric Clustering, which constructs a sparse anisotropic topology to represent non-reciprocal predictive dependencies beyond symmetric similarity. Extensive experiments on diverse and challenging real-world benchmarks demonstrate that TVTS consistently achieves state-of-the-art forecasting performance with high computational efficiency across forecasting horizons.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series Forecasting, Time-Varying, Cross-Variable Dependencies

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 4 reviews, above 85% of the others.
Reviewer_frFPOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TVTS, a multivariate time series forecasting framework that jointly models time-varying temporal dynamics and directed cross-variable dependencies. The method consists of a Time-Varying Convolutional Filtering module and a Directed Geometric Clustering module, and combines topology-guided attention with frequency-domain residual prediction to produce the final forecasts.

The authors conduct relatively extensive experiments on long-term, short-term, and several additional forecasting benchmarks, reporting good forecasting accuracy and computational efficiency.

Critical strengths

S1: The motivation of the paper is relatively clear. It considers both time-varying patterns and directional cross-variable dependencies, and the overall framework is relatively complete.

S2: The experimental coverage is relatively broad, including recent competitive baselines, as well as ablation, stability, and computational efficiency analyses.

S3: The model achieves good forecasting performance, particularly showing relatively significant advantages on the PEMS datasets.

Critical weaknesses

W1. The novelty of the core method is incremental.

TVCF mainly implements time-varying filtering through position-dependent coefficients and shared basis expansion, while DGC constructs an asymmetric adjacency matrix through independent source and target embeddings and Top-k selection. These core designs all have corresponding technical foundations in existing research.

W2. The independent benefits of directed dependency modeling are limited, and ablation validation on key high-dimensional datasets is insufficient.

Although Table 4 provides an ablation comparison with bidirectional topology (Dual-A), the performance differences between the two designs are small on most datasets, which is insufficient to fully demonstrate the necessity of directed topology. In addition, the ablation experiments lack results on high-dimensional traffic datasets such as Traffic and PEMS. Considering that directed propagation among traffic sensors is one of the main research motivations of this paper, and that the model achieves relatively significant performance advantages on the PEMS datasets, the existing experiments are insufficient to verify the independent contribution of directed dependency modeling in these scenarios.

W3. Some baseline results are inconsistent with the metrics reported in existing literature, and the experimental configurations and result sources are not sufficiently transparent.

The experimental results of some baseline models differ from the metrics reported in their original papers or official implementations, but the paper does not sufficiently explain the sources of these differences.

W4. The data source and illustrative nature of Figure 1 are unclear.

Figure 1 shows directed propagation paths among traffic sensors, but it does not clearly indicate whether the figure is based on real traffic data and sensor topology or is merely a conceptual illustration used to explain the research motivation. If it is based on real data, the data source and the basis for determining propagation directions should be explained. If it is only an illustration, this should be clearly stated in the figure caption to avoid interpreting hypothetical propagation relationships as empirical observations.

Questions for authors

Q1: Which designs in TVCF and DGC constitute substantive innovations compared with existing methods? Can more direct comparative experiments be provided to verify the independent contributions of these designs?

Q2: How is Dual-A in Table 4 constructed? Why are the performance differences between directed and bidirectional topologies small on most datasets? Can ablation experiments on the Traffic and PEMS datasets be added, including controlled comparisons among directed topology, symmetric topology, and standard full Attention?

Q3: DGC constructs an asymmetric adjacency matrix through two sets of static node embeddings. How do the authors verify that the learned asymmetry corresponds to real directional predictive dependencies, rather than merely increasing the flexibility of relational representations? Can quantitative validation be provided on data with known directional relationships?

Q4: The position-dependent coefficients (C) in TVCF are globally learnable parameters and do not vary with input samples. How do the authors explain the adaptability of this design to non-stationary dynamics across different samples? Compared with standard convolution and directly learning position-dependent convolution kernels, how much additional benefit does structured basis expansion provide? Can corresponding comparative experiments be added?

Q5: For baselines whose reported metrics differ from those in the original papers or official implementations, can the authors explicitly list the result sources, training and evaluation configurations, and specific adjustments made relative to the official settings, and explain the performance differences caused by these adjustments?

Q6: Is Figure 1 drawn based on real traffic sensor data, or is it merely a conceptual illustration? If it is based on real data, how are the propagation paths and directions determined?

6 comments

We sincerely thank the reviewer for the thoughtful assessment and constructive feedback. We address each concern through methodological clarifications and additional controlled experiments supporting the proposed design.

Q1 / Weakness 1: Methodological Novelty and Component Contributions

The methodological contribution centers on unified temporal-relational forecasting. TVCF builds upon node-variant graph filtering [G1,G2] and the normalized complex basis parameterization of TV-Rec [G3]. Similarly, source-target embeddings and graph selection have established foundations in Graph WaveNet [G4] and MTGNN [G5]. We acknowledge these connections and will further clarify the technical distinctions and attribution.

TVTS jointly models complementary temporal and relational dependencies for multivariate forecasting. TVCF captures position-dependent temporal dynamics, while DGC learns non-reciprocal predictive relationships across variables. These modules characterize distinct temporal and cross-variable dimensions. TGF incorporates DGC-guided interactions into TVCF-filtered representations, followed by FDRP for spectral residual refinement in multi-step forecasting. Their coordinated design enables the joint characterization of evolving temporal patterns and directional cross-variable interactions within a unified framework.

Controlled experiments substantiate the contributions of the proposed components. We additionally compare temporal-filter alternatives (static convolutional banks, positional gating, and direct position-specific kernels) and graph affinity learners (Graph WaveNet-style and MTGNN-style) under the same TVTS backbone. On ECL, TVCF achieves an MSE of 0.157 versus 0.161–0.162 for temporal alternatives, while DGC achieves 0.157 versus 0.164/0.168 for graph alternatives. Table 4 further shows that removing FDRP, TVCF, or DGC increases ETTm1 MSE from 0.375 to 0.378, 0.380, and 0.378, respectively. Table 20 also reports lower MSE for TVTS (0.375) than DUET (0.390) and BIM$^3$ (0.387). Collectively, these comparisons provide empirical support for the effectiveness of each component and their integration.

Q2 / Weakness 2: Dual-A Construction and Traffic Ablations

Dual-A adopts DUET-style metric-based relational modeling, distinct from explicit DGC symmetrization. Following DUET, Dual-A computes channel similarities from Fourier magnitudes using a learned Mahalanobis-type distance to construct cross-variable connectivity. (Implementation: https://github.com/decisionintelligence/DUET/blob/main/ts_benchmark/baselines/duet/utils/masked_attention.py.) While the underlying distance is symmetric, subsequent row normalization and mask sampling can produce asymmetric effective interactions. We will clarify this construction and its distinction from DGC. Additionally, a standard "full attention" mechanism is included as a baseline.

Additional traffic ablations demonstrate the predictive advantages of DGC. We extend the comparisons to Traffic and PEMS03/04/07/08 under the same TVTS backbone and training protocol. Table R1 presents single-seed results averaged over H=96,192,336,720H=96,192,336,720 for Traffic and H=12,24,48,96H=12,24,48,96 for PEMS (L=96L=96). TVTS consistently achieves lower MSE and MAE than Dual-A and full attention across all five datasets. Compared with Dual-A, DGC reduces MSE by 11.8%, 14.5%, and 17.6% on PEMS03, PEMS04, and PEMS07, respectively. These improvements are more pronounced than those on several ETT datasets in Table 4, highlighting the value of structured relational modeling in traffic forecasting.

The comparisons highlight the benefits of learned relational modulation and structured affinity modeling. The performance gap relative to full attention demonstrates the predictive utility of incorporating relational information. Dual-A further provides a competitive metric-based relational baseline, differing from DGC in affinity construction and reciprocity assumptions. Accordingly, the observed improvements characterize the effectiveness of the overall relational design. We interpret these single-seed results as additional predictive evidence, while statistical significance and physical directional alignment require complementary validation.

Table R1. Forecasting performance under different relational mechanisms on traffic datasets (L=96L=96).

DatasetTVTS (DGC) MSETVTS (DGC) MAEDual-A MSEDual-A MAEFull Attention MSEFull Attention MAE
Traffic0.4280.2770.4410.2940.4430.301
PEMS030.1200.2260.1360.2460.1370.244
PEMS040.1120.2150.1310.2330.1340.236
PEMS070.1030.2030.1250.2290.1280.231
PEMS080.1450.2290.1650.2520.1680.259

Due to space constraints in individual response boxes, we will address the remaining questions separately in subsequent responses.

Q3: Quantitative Validation of Directed Dependencies

Directional validity requires evidence beyond learned asymmetry. Figures 9–11 visualize asymmetric relations, while Appendix J.2.2 and Figure 12 assess their correspondence with Granger-based predictive directions. To further strengthen directional validation, we introduce controlled synthetic systems with known dependencies and complementary real-world experiments with physically grounded reference directions.

Controlled synthetic systems enable quantitative evaluation of directional recovery. We construct 16-variable VAR(1) processes with randomly relabeled chain and branching-tree topologies:

\quad \mathbf{e}(t) \sim \mathcal{N}(0, 0.25\mathbf{I}).$$ Self-couplings and directed cross-variable couplings are set to 0.45 and 0.35, respectively, yielding a spectral radius of 0.45. Innovations are independent across variables and time, and the generating matrix $\mathbf{A}$ specifies the ground-truth directed edges. Table R2 evaluates constructed relation scores using direction agreement (DA), tie rate, AUROC, average precision (AP), and precision at 15 (P@15). AUROC and AP cover all 240 off-diagonal ordered pairs, including 15 true directed edges (6.25% prevalence). **Illustrative asymmetric scores exhibit strong directional correspondence across both structures.** On the chain, DGC-labeled scores achieve 1.00 DA, 0.89 AUROC, 0.53 AP, and 0.73 P@15, outperforming shuffled scores (0.53/0.51/0.07/0.00). On the branching structure, the corresponding results are 0.93/0.86/0.37/0.67 versus 0.60/0.56/0.13/0.13. Symmetric scores show weaker edge-ranking performance, with zero DA reflecting directional ties by construction. These comparisons illustrate how asymmetric representations can encode directional preferences and identify ground-truth edges under different dependency structures. **Synthetic analysis provides a controlled directional evaluation protocol.** The constructed scores demonstrate how directional agreement and edge recovery can be quantified, with P@15 reflecting partial recovery under both structures. Since these illustrative scores are constructed rather than extracted from forecasting-trained DGC, we further assess actual learned relations on real-world data with independently established physical directions. **Table R2.** Illustrative directional recovery on synthetic scores ($L=96$, $H=12$). |System|Method|DA|Ties|AUROC|AP|P@15| |:---|:---|---:|---:|---:|---:|---:| |Chain|DGC (weights)|1.00|0.00|0.89|0.53|0.73| ||DGC (affinity)|0.67|0.00|0.68|0.58|0.67| ||Shuffled weights|0.53|0.00|0.51|0.07|0.00| ||Symmetric (weights)|0.00|1.00|0.84|0.33|0.40| ||Symmetric (affinity)|0.00|1.00|0.79|0.29|0.40| ||OLS|1.00|0.00|1.00|1.00|1.00| |Branching|DGC (weights)|0.93|0.00|0.86|0.37|0.67| ||DGC (affinity)|0.67|0.00|0.60|0.27|0.60| ||Shuffled weights|0.60|0.00|0.56|0.13|0.13| ||Symmetric (weights)|0.00|1.00|0.52|0.06|0.00| ||Symmetric (affinity)|0.00|1.00|0.45|0.07|0.07| ||OLS|1.00|0.00|1.00|1.00|1.00| **Additional Real-World Validation of Directed Dependencies** **Forecasting-trained DGC captures partial alignment with physical directions in real-world river networks.** We further evaluate DGC on CausalRivers [G6], which provides measured river discharge and known upstream-to-downstream directions. Appendix J.4 and Figure 15 present two cases: the Oder watershed (22 gauges, 21 directed edges) and an Elbe flood event (42 gauges, 42 directed edges). TVTS is trained solely with forecasting supervision, using the physical topology only for evaluation. For each reference edge $u\rightarrow v$, its orientation is correctly identified when the learned weights satisfy $W_{vu}>W_{uv}$. **Real-world results provide complementary evidence of directional predictive relevance.** DGC correctly identifies 13/21 reference directions (61.9%) on Oder and 26/42 (61.9%) on Elbe. These findings directly evaluate forecasting-trained relational weights and demonstrate partial correspondence with externally grounded physical directions. Together with the Granger-based analysis, these results support DGC's ability to learn directionally informative predictive relations, while full causal topology recovery remains beyond the scope of this study.

Q4: Position Dependence and Structured Basis Expansion

TVCF models position-dependent temporal dynamics through globally shared filters. In Eq. (5), shared coefficients C\mathbf{C} and normalized spectral bases Bˉ\bar{\mathbf{B}} define position-specific filters:

ht,q=∑m=1MCt,mBˉm,q.h_{t,q}=\sum_{m=1}^{M}C_{t,m}\bar{B}_{m,q}.

Unlike shift-invariant convolution, TVCF assigns position-specific lag responses to capture recurring within-window dynamics. While C\mathbf{C} is shared across samples, input-dependent normalization and content attention capture sample-specific variations. Together, these mechanisms enable position-dependent modeling and input-adaptive processing.

Random initialization provides a competitive default for shared filter coefficients. We compare context-aware and random initialization of C\mathbf{C} under identical architecture and optimization. Table R3 averages H∈[96,192,336,720]H\in[96,192,336,720] for long-term datasets and H∈[12,24,48,96]H\in[12,24,48,96] for PEMS (L=96L=96). Following the relative ADF statistics in TimeBridge [G7], we highlight ETTh1, ETTh2, ETTm2, and ECL, whose less negative values suggest comparatively stronger non-stationary tendencies under this diagnostic.

Initialization strategies exhibit dataset-dependent effectiveness. Random initialization matches or slightly improves upon context-aware initialization on the highlighted datasets (ECL MSE: 0.157 vs. 0.158; ETTm2: 0.270 vs. 0.271; ETTh2: 0.363 vs. 0.365). A possible explanation is that input-derived initialization introduces transient fluctuations into shared coefficients. Conversely, context-aware initialization achieves modest MSE improvements on PEMS03 (0.119 vs. 0.120), PEMS07 (0.102 vs. 0.103), and PEMS08 (0.144 vs. 0.145), potentially benefiting from recurring temporal patterns. These findings suggest that contextual initialization can provide useful prior information under favorable conditions, while random initialization remains a robust default. The experiment evaluates initialization sensitivity, with small differences requiring further statistical validation.

Table R3. Context-aware versus random initialization of C\mathbf{C} (L=96L=96).

DatasetContext-aware MSEContext-aware MAERandom (TVTS) MSERandom (TVTS) MAE
ETTh10.4280.4270.4280.426
ETTh20.3650.3910.3630.390
ETTm10.3740.3850.3750.386
ETTm20.2710.3150.2700.315
Weather0.2410.2680.2400.267
Exchange0.3330.3890.3340.390
Traffic0.4290.2770.4280.277
ECL0.1580.2490.1570.249
PEMS030.1190.2260.1200.226
PEMS040.1120.2150.1120.215
PEMS070.1020.2040.1030.203
PEMS080.1440.2290.1450.229

Structured TVCF offers advantages over conventional position-aware filtering. We compare TVCF with a position-shared nonlinear convolutional bank (Static Bank), its positional-gating extension (Gated Bank), and directly learned position-specific real-valued kernels (Direct Taps). All share the backbone, temporal support, normalization, and training budget. Bank widths target comparable filter-parameter budgets, while Direct Taps has fewer parameters. Basis orthogonality regularization applies only to TVCF.

Controlled comparisons demonstrate the benefits of structured TVCF. Table R4 shows that TVCF achieves the lowest averaged MSE and MAE across all 12 datasets. On Traffic, its MSE is 0.428 versus 0.441/0.445/0.443 for Static Bank/Gated Bank/Direct Taps; on ECL, the results are 0.157 versus 0.161/0.162/0.161. These improvements demonstrate the effectiveness of TVCF's structured position-dependent parameterization beyond positional gating or unconstrained kernels. Given differences in parameter capacity and regularization, these comparisons substantiate the overall TVCF design, encompassing spectral structure and regularization.

Table R4. Temporal filtering alternatives within TVTS (L=96L=96).

DatasetStatic Bank MSEStatic Bank MAEGated Bank MSEGated Bank MAEDirect Taps MSEDirect Taps MAETVCF MSETVCF MAE
ETTh10.4340.4300.4470.4360.4350.4300.4280.426
ETTh20.3660.3940.3680.3950.3670.3930.3630.390
ETTm10.3790.3890.3800.3890.3780.3890.3750.386
ETTm20.2740.3190.2750.3180.2740.3180.2700.315
Weather0.2450.2680.2470.2700.2460.2690.2400.267
Exchange0.3360.3930.3380.3940.3370.3920.3340.390
Traffic0.4410.2810.4450.2830.4430.2830.4280.277
ECL0.1610.2540.1620.2540.1610.2530.1570.249
PEMS030.1240.2300.1260.2290.1260.2300.1200.226
PEMS040.1180.2180.1200.2200.1160.2190.1120.215
PEMS070.1080.2050.1100.2050.1060.2050.1030.203
PEMS080.1470.2320.1480.2330.1470.2330.1450.229
Reviewer_bc6gOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TVTS, a multivariate time series forecasting framework designed to model position-dependent temporal dynamics and directed cross-variable dependencies. TVTS consists of four main components: (1) a Time-Varying Convolutional Filter (TVCF) that models sequences on a zero-padded directed cyclic graph and parameterizes position-specific spectral filter taps via a low-rank complex basis expansion; (2) Directed Geometric Clustering (DGC), which learns decoupled static source and target channel embeddings to construct a sparse, asymmetric Top-k attention mask; (3) Topology-Guided Fusion (TGF), which injects this directed mask into cross-variable attention; and (4) Frequency-Domain Residual Prediction (FDRP), which stacks complex-valued spectral MLPs to map the representation to the forecasting horizon. Experiments across 17 benchmarks (including long-term, short-term PEMS, M4 univariate, and few-shot settings) show strong empirical results.

Critical strengths
  1. The motivation of relaxing shift-invariant temporal filtering and symmetric channel similarity in multivariate time series forecasting is intuitive. Formulating causal linear convolution via zero-padded circulant shift operators and combining it with sparse directed channel attention and frequency-domain residual blocks yields a unified, lightweight architecture.

  2. The empirical evaluation is extensive in breadth, covering 17 datasets across long-term forecasting, short-term traffic forecasting (PEMS), univariate forecasting (M4), and few-shot regimes, accompanied by runtime and memory profiling.

Critical weaknesses

The empirical results in the ablation and permutation studies do not convincingly support the paper's central claims regarding directed topology learning (DGC) and time-varying filtering (TVCF).

First, in Table 4 (and Table 17), replacing the directed topology with a symmetric bidirectional topology (Dual-A) yields almost identical performance to the full TVTS model across nearly all datasets (e.g., 0.270 vs. 0.270 on ETTm2, 0.429 vs. 0.428 on ETTh1, 0.241 vs. 0.240 on Weather, and 0.376 vs. 0.375 on ETTm1), and replacing the time-varying filter with a static filter (Based-H) also results in tiny differences of 0.001 to 0.003 in MSE. Second, in Appendix K (Table 18), the authors report that randomly permuting 100% of the input variable order at inference time (X' = PX) leaves the test MSE virtually unchanged (e.g., 0.157 vs. 0.158 on ECL with 321 variables). Because the DGC source and target embeddings in Eq. (6) are static learnable parameter tables indexed by channel ID rather than computed dynamically from the input series, permuting the input channels at test time mismatches every variable with the wrong directed edges in M. If scrambling the learned directed graph at inference time causes virtually no degradation, it suggests that the forecasting accuracy is driven primarily by the channel-independent spectral blocks rather than the learned directed topology.

Questions for authors
  1. In Appendix K (Table 18), since the source and target embeddings in Eq. (6) are static parameter matrices tied to channel indices, permuting 100% of the input channels in X at inference time connects each variable to mismatched neighbors in the mask M. Why does test MSE on ECL remain unchanged (0.157 to 0.158) when the directed graph is scrambled at inference time, whereas replacing M with Random-M in Table 4 degrades ECL MSE from 0.157 to 0.178?

  2. In Table 4, the symmetric variant Dual-A and the static filter variant Based-H perform within 0.001 to 0.003 MSE of the full TVTS model on most datasets. Given that the standard deviations across random seeds are 0.001 to 0.004 (Table 16), are the improvements of TVTS over Dual-A and Based-H statistically significant?

2 comments

We sincerely thank the reviewer for the careful assessment and constructive feedback. We address each concern in turn, providing additional experimental evidence and clarifications.

Q1 / Weakness 1: Permutation Sensitivity and Directed Topology

Channel-order robustness and sensitivity to learned relations probe distinct properties. Appendix K (Table 18) reports a small performance change under its original channel-permutation protocol (ECL MSE: 0.157 to 0.158). To further examine this behavior, we introduce explicit controls that distinguish consistent channel reordering from graph-variable misalignment. DGC learns a dataset-level asymmetric predictive prior, where historical observations from one variable may inform another's future state (e.g., congestion at sensor B preceding that at sensor A). These relations represent directional predictive relevance rather than physical causal paths, and their effectiveness depends on preserving variable-relation correspondence.

Controlled relabeling and topology interventions reveal complementary behaviors. We conduct fixed-checkpoint experiments without retraining (Table R1), averaging four prediction horizons (H∈[96,192,336,720]H\in[96,192,336,720]). Consistent channel relabeling jointly permutes inputs, targets, graph indices, and RevIN affine parameters, preserving their correspondence. ECL MSE changes only from 0.157 to 0.159 (Exchange: 0.334 to 0.336). In contrast, permuting inputs while keeping the graph and affine parameters fixed increases ECL MSE to 0.330; aligning affine parameters reduces it to 0.203. More directly, graph-index permutation and within-row shuffling increase MSE to 0.204 and 0.197, respectively. The latter preserves diagonal entries and row-wise weight distributions, highlighting the importance of relational assignments. Exchange exhibits smaller but consistent degradation (0.345–0.354), indicating dataset-dependent sensitivity.

Random-M evaluates relational learning rather than inference-time permutation robustness. Unlike fixed-checkpoint interventions, Random-M replaces DGC with a randomly initialized, trainable relation matrix throughout training, changing the relational inductive bias and optimization process. Although trainable, this alternative lacks DGC's structured source-target parameterization and may capture informative cross-variable interactions less effectively. Less selective relational weighting may also introduce interference from irrelevant channels, consistent with the motivation of selective channel modeling. The resulting ECL MSE increase (0.157 to 0.178) supports the utility of DGC's relational learning mechanism.

Together, these experiments distinguish permutation consistency from reliance on learned topology. Consistent relabeling largely preserves predictive performance, whereas graph-index permutation and row-wise shuffling disrupt learned relational assignments and degrade forecasting accuracy. These results demonstrate the predictive contribution of learned relations, with DGC serving as a directional predictive prior rather than a physical causal topology estimator.

Table R1. Fixed-checkpoint inference interventions on ECL and Exchange (L=96L=96). No retraining is performed.

InterventionECL MSEECL MAEExchange MSEExchange MAE
Original0.1570.2490.3340.390
Consistent channel relabeling0.1590.2500.3360.394
Input perm. (graph/affine fixed)0.3300.3530.3480.408
Input perm. (affine aligned)0.2030.2810.3460.403
Graph-index perm.0.2040.2840.3450.405
Row-wise shuffle (diagonal fixed)0.1970.2760.3460.405
Uniform weights (1/N1/N)0.1850.2730.3490.408
Frozen Gaussian weights0.2570.3360.3540.411

Due to space constraints in individual response boxes, we will address the remaining questions separately in the subsequent responses.

Q2 / Weakness 2: Statistical Significance of Ablation Improvements

Existing significance tests do not directly assess improvements over Dual-A and Based-H. Table 16 evaluates statistical significance against external forecasting baselines rather than these ablation variants. Table 4 also shows dataset-dependent gains: TVTS achieves MSEs of 0.157/0.334 on ECL/Exchange, compared with 0.162/0.355 for Dual-A and 0.161/0.352 for Based-H, while differences on several ETT and Weather datasets are smaller.

Five-seed experiments demonstrate consistent average improvements. Table R2 reports mean ±\pm standard deviation over five random seeds, with each seed's results averaged across H∈[96,192,336,720]H\in[96,192,336,720] (L=96L=96). Based-H replaces TVCF with position-shared filtering, while Dual-A replaces the learned asymmetric topology with a bidirectional counterpart. TVTS achieves lower mean MSE than Based-H on all seven datasets and Dual-A on six, with a rounded tie on ETTm2. Its mean MAE is lower than both variants across all seven datasets. The gains are particularly pronounced on Exchange, with MSE reductions of 0.018 and 0.021 over Based-H and Dual-A, respectively, whereas several ETT and Weather improvements are more modest.

Statistical interpretation requires paired comparisons across seeds. The five-seed results show consistent average gains, particularly on ECL and Exchange. However, marginal standard deviations alone cannot establish the significance of paired performance differences. A formal assessment requires paired per-seed tests, followed by multiple-comparison correction. We therefore present Table R2 as repeated-seed evidence of consistent improvements, while avoiding claims of statistical significance across all datasets.

Table R2. Five-seed ablation results (mean ±\pm standard deviation, L=96L=96).

DatasetTVTS MSETVTS MAEBased-H MSEBased-H MAEDual-A MSEDual-A MAE
ETTm10.375±0.0010.375\pm0.0010.386±0.0010.386\pm0.0010.377±0.0020.377\pm0.0020.388±0.0010.388\pm0.0010.376±0.0020.376\pm0.0020.391±0.0020.391\pm0.002
ETTm20.270±0.0010.270\pm0.0010.315±0.0010.315\pm0.0010.272±0.0010.272\pm0.0010.316±0.0020.316\pm0.0020.270±0.0020.270\pm0.0020.316±0.0010.316\pm0.001
ETTh10.428±0.0010.428\pm0.0010.426±0.0010.426\pm0.0010.433±0.0020.433\pm0.0020.427±0.0030.427\pm0.0030.429±0.0020.429\pm0.0020.427±0.0020.427\pm0.002
ETTh20.363±0.0020.363\pm0.0020.390±0.0010.390\pm0.0010.366±0.0030.366\pm0.0030.392±0.0020.392\pm0.0020.366±0.0030.366\pm0.0030.391±0.0020.391\pm0.002
ECL0.157±0.0010.157\pm0.0010.249±0.0010.249\pm0.0010.161±0.0020.161\pm0.0020.254±0.0020.254\pm0.0020.162±0.0020.162\pm0.0020.253±0.0020.253\pm0.002
Exchange0.334±0.0020.334\pm0.0020.390±0.0010.390\pm0.0010.352±0.0040.352\pm0.0040.401±0.0030.401\pm0.0030.355±0.0030.355\pm0.0030.399±0.0030.399\pm0.003
Weather0.240±0.0040.240\pm0.0040.267±0.0030.267\pm0.0030.244±0.0060.244\pm0.0060.270±0.0030.270\pm0.0030.241±0.0050.241\pm0.0050.269±0.0040.269\pm0.004

We sincerely appreciate the reviewer's thoughtful feedback. We hope these additional experiments and clarifications address the concerns raised.

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

This paper proposes TVTS, a unified framework for multivariate time series forecasting that aims to address two limitations of existing approaches: the inability of time-invariant temporal filters to capture position-dependent temporal dynamics, and the reliance on symmetric similarity-based relationships to characterize cross-variable dependencies. Experiments on 17 real-world datasets show consistent state-of-the-art performance.

Critical strengths
  1. The paper identifies two important challenges: temporal heterogeneity and asymmetric cross-variable dependencies. These are relevant issues in real-world multivariate systems, particularly traffic networks and other systems where information propagation is directional.
  2. The TVCF module is derived from node-variant graph filters with a clean spectral equivalence proof.
  3. TVTS achieves the best results on 15 of the 16 aggregated MSE/MAE comparisons in the long-term forecasting.
Critical weaknesses
  1. The novelty of the proposed TVCF mechanism needs to be better distinguished from existing node-variant graph filtering approaches. The core idea of TVCF is to assign different filter coefficients to different temporal positions, thereby relaxing the shift-invariance assumption of conventional convolution. However, the mathematical formulation is closely related to node-variant graph filters, which also assign node-specific filtering coefficients. Indeed, the paper explicitly builds upon this line of research and derives its spectral representation using existing graph filtering formulations. The relationship to existing time-varying filtering approaches, particularly TV-Rec [1], also deserves a more explicit discussion.
  2. DGC constructs an asymmetric adjacency matrix using separate source and target embeddings. Although this formulation can produce asymmetric relationships, asymmetry alone does not establish that the learned edges correspond to meaningful directional predictive dependencies. In particular, the source and target embeddings appear to be freely learned parameters, without explicit temporal precedence constraints. Consequently, the learned adjacency may simply represent an asymmetric latent interaction pattern optimized for the forecasting objective, rather than genuine directional information propagation.
  3. Equation (12) defines the orthogonal regularization. However, when M>K+1, the rank of B~pB~p⊤\widetilde B_p\widetilde B_p^\top cannot exceed K+1. Therefore, it is mathematically impossible for this matrix to equal the M×MM\times M identity matrix.
  4. The authors extend the original sequence by appending K-1 zero-valued observations to construct a directed cyclic graph. The motivation is to suppress circular boundary artifacts introduced by cyclic shifts. However, the paper does not sufficiently establish that this padding strategy completely eliminates unwanted cyclic interactions under the proposed time-varying filtering operation.
  5. Table 7 reveals substantial dataset-specific tuning. The sensitivity analysis (Figures 5, 15) shows that performance is sensitive to M, k, and β\beta. This raises concerns about how well TVTS generalizes to new datasets without extensive hyperparameter search.

[1] TV-Rec: Time-variant convolutional filter for sequential recommendation. NeurIPS, 2025.

Questions for authors

Please refer to Critical Weaknesses.

9 comments

We sincerely thank the reviewer for the constructive feedback. We clarify our contributions and provide additional experimental evidence below. Tables R1, R2, etc. refer to rebuttal results, while Tables 1, 2, etc. refer to the revised manuscript.

Q1 / Weakness 1: Relationship to Node-Variant Filtering and TV-Rec

TVTS learns which filtered histories should inform each variable's forecast. TVCF follows node-variant graph filtering [G1, G2] in allowing the same lag to receive different weights at different temporal positions. It adopts TV-Rec’s spectral formulation, in which each position combines a shared set of normalized complex filters [G3]. TV-Rec uses a user’s filtered item history to rank candidates for the next interaction. TVTS filters each variable’s history separately and learns a second graph over variables. On ECL, for example, this graph allows one electricity client’s demand history to inform another client’s forecast even when the reverse connection is absent. The graph selects source variables for each target, and attention weights their contributions for the current input window.

TVTS couples temporal filtering with directed cross-variable aggregation. The temporal graph underlying TVCF connects positions within a history window, whereas DGC learns relations among distinct observed variables. These mechanisms capture complementary dependencies: TVCF models position-dependent temporal patterns within each variable, while DGC provides asymmetric predictive priors for cross-variable interactions. Specifically, TGF integrates DGC-guided interactions over TVCF-filtered representations, followed by FDRP for multi-step spectral residual prediction. Thus, TVTS extends beyond adapting the forecasting objective by explicitly integrating temporal filtering with directed relational modeling.

Component ablations support the contributions across datasets. Across all seven datasets in Table 4, removing TVCF, DGC, or FDRP increases both horizon-averaged MSE and MAE. On ECL, MSE rises from 0.157 for TVTS to 0.172 without TVCF, 0.184 without DGC, and 0.160 without FDRP. These results demonstrate the complementary contributions of temporal filtering, relational modeling, and spectral residual prediction within the unified architecture, with consistent improvements across datasets.

Topology replacements further validate the relational design. With TVCF retained, replacing topology-guided aggregation with Full Attention, Random-M, or Dual-A increases ECL MSE from 0.157 to 0.175, 0.178, and 0.162, respectively. Across seven datasets, TVTS achieves lower MSE and MAE than both Full Attention and Random-M, while outperforming Dual-A in 13 of 14 comparisons, with one MSE tie. These findings support the predictive effectiveness of learned relational structure beyond simply introducing cross-variable attention.

Due to space constraints in individual response boxes, we will address the remaining questions separately in the subsequent responses.

Q2 / Weakness 2: Interpretation and Evidence for Directed Relations

DGC captures directional predictive dependencies through asymmetric relational modeling. The learned source-target embeddings define a dataset-level asymmetric relational prior shared across input windows. This topology captures persistent directional predictive relevance without explicit temporal-precedence constraints, while input-dependent attention dynamically adjusts cross-variable interactions. Accordingly, the learned edges characterize how information from one variable contributes to forecasting another.

Quantitative evaluation reveals dataset-dependent alignment with predictive lead-lag relationships. Figure 11 visualizes asymmetric topologies across eight datasets, including PEMS04 and PEMS08. To further assess directional correspondence, Table R2 extends the Granger-based analysis in Appendix J.2.2 by comparing learned affinities with predictive references across four lag orders. We report direction agreement (DA, %) and Spearman correlation (ρ\rho), counting ties as nonagreement. DGC achieves 59.2–68.3% DA on ECL and 55.0–60.0% on PEMS07, with positive Spearman correlations across all evaluated lags. Weather and Traffic exhibit more variable alignment, reflecting dataset-dependent directional patterns. Overall, these results demonstrate measurable correspondence between learned relations and predictive lead-lag structure.

Controlled comparisons establish the predictive effectiveness of DGC. Table R1 compares DGC with Graph WaveNet-style (GWNet) [G4] and MTGNN-style [G5] affinity learners under the same TVTS backbone, Top-kk budget, and fusion mechanism, with comparable graph-specific parameter counts (PgP_g). Results are averaged across four prediction horizons per dataset. DGC achieves the lowest MSE and MAE on all 12 datasets, including ECL (MSE: 0.157 vs. 0.164/0.168) and PEMS07 (0.103 vs. 0.116/0.115) compared with GWNet/MTGNN. These improvements demonstrate the predictive value of DGC's structured asymmetric relational modeling.

For clarity, Tables R1–R2 report additional rebuttal experiments, while Figure 11 and Appendix J.2.2 refer to the revised manuscript.

Table R1. Forecasting with alternative affinity learners in TVTS (L=96L=96; PgP_g: graph-specific parameters).

DatasetGWNet PgP_gGWNet MSEGWNet MAEMTGNN PgP_gMTGNN MSEMTGNN MAEDGC PgP_gDGC MSEDGC MAE
ETTh12240.4370.4292100.4430.4332240.4280.426
ETTh22240.3720.3952100.3740.3962240.3630.390
ETTm12240.3790.3912100.3810.3922240.3750.386
ETTm22240.2750.3192100.2740.3182240.2700.315
Weather6720.2510.2726400.2560.2756720.2400.267
Exchange2560.3390.3942720.3410.3952560.3340.390
Traffic275840.4350.281281280.4380.282275840.4280.277
ECL102720.1640.255101100.1680.257102720.1570.249
PEMS03114560.1230.228112200.1280.233114560.1200.226
PEMS0498240.1130.21696900.1340.22698240.1120.215
PEMS07282560.1160.207288000.1150.207282560.1030.203
PEMS0854400.1520.23353380.1610.24154400.1450.229

Table R2. Directional correspondence with Granger references (L=H=96L=H=96; DA in %, ρ\rho: Spearman correlation).

DatasetMethodp=3p=3 DAp=3p=3 ρ\rhop=6p=6 DAp=6p=6 ρ\rhop=12p=12 DAp=12p=12 ρ\rhop=24p=24 DAp=24p=24 ρ\rho
WeatherGWNet27.5-0.09326.7-0.18826.7-0.19829.2-0.156
MTGNN46.70.10450.50.12048.30.11949.20.107
Random47.5-0.13548.3-0.11142.5-0.14043.3-0.146
DGC55.80.01950.00.09947.50.15146.70.158
TrafficGWNet39.20.09943.30.12542.50.11544.20.150
MTGNN44.2-0.10243.3-0.11342.5-0.05252.50.013
Random45.80.05851.70.06249.20.01850.80.052
DGC52.00.12254.50.11153.00.14556.70.160
ECLGWNet39.20.14838.30.13336.70.10741.70.173
MTGNN45.8-0.07245.8-0.03950.00.01046.70.021
Random46.7-0.07248.3-0.06850.8-0.04154.2-0.066
DGC59.20.13965.80.18668.30.22465.00.273
PEMS07GWNet35.8-0.05637.5-0.06334.2-0.10138.3-0.084
MTGNN50.00.07652.50.07050.80.02950.00.032
Random54.2-0.09051.7-0.05848.3-0.06345.8-0.074
DGC60.00.18957.50.14655.80.11055.00.139

Q3 / Weakness 3: Overcomplete Bases and Orthogonal Regularization

The regularizer encourages basis diversity rather than exact orthogonality. We agree that when M>d=K+1M > d = K+1, the Gram matrix cannot equal the identity matrix IMI_M due to its rank constraint. For either row-normalized real or imaginary component, let U∈RM×dU \in \mathbb{R}^{M \times d} have unit-norm rows. Then,

= \sum_{i \ne j}(u_i^\top u_j)^2 \ge \frac{M^2}{d} - M > 0.$$ The lower bound follows from $\operatorname{tr}(UU^\top)=M$ and $\operatorname{rank}(UU^\top) \le d$. Therefore, Eq. (12) penalizes pairwise basis correlations but cannot achieve zero regularization loss in the overcomplete regime. We will revise the description from "enforcing orthogonality" to "encouraging basis diversity." The ablation in Table 4 supports its empirical utility: removing the regularizer increases ECL MSE from 0.157 to 0.162, without implying that exact orthogonality is attainable. `Q4 / Weakness 4: Zero Padding and Circular Boundaries` **Complete suppression of circular wrap-around requires at least $K$ trailing zeros.** For shift orders $q=0,\ldots,K$, let $x_{\mathrm{pad}}=[x;0_P]$ with $P \ge K$, and let $S_{\mathrm{cyc}}$ denote the cyclic shift operator. Define the filtered sequence as $$y = \sum_{q=0}^{K} \operatorname{diag}(w_q) S_{\mathrm{cyc}}^q x_{\mathrm{pad}}.$$ For each retained position $1 \le t \le L$, the output satisfies $$y_t = \sum_{q=0}^{\min(K,t-1)} w_{t,q}x_{t-q}.$$ Thus, all wrapped contributions originate from the zero buffer, regardless of the position-dependent filter coefficients. **The current $K-1$ padding leaves one boundary case unresolved.** Specifically, the $q=K$ shift can wrap the final historical observation into the first retained position. We will correct the padding condition in Appendix B and revise the claim of complete artifact elimination accordingly. This boundary effect involves interactions within the observed history, rather than future-target leakage. All reported experiments use the same boundary convention; the performance impact of corrected padding remains to be evaluated. `Q5 / Weakness 5: Hyperparameter Tuning and Generalization` **TVTS adopts dataset-specific hyperparameter configurations.** As documented in Table 7 and Appendix G.3, several parameters are selected individually, while related datasets share certain settings. For example, all four ETT datasets use Top-$k=2$, whereas ECL and Traffic use 10 and 25, respectively. These configurations provide practical starting points for related datasets. **Targeted validation provides a practical strategy for unseen datasets.** We recommend initializing from the closest documented configuration and tuning the key parameters $M$, $k$, and $\beta$ on validation data, rather than repeating the entire search. Figures 5 and 15 illustrate parameter sensitivity and guide this selection. Accordingly, the cross-dataset results demonstrate forecasting effectiveness under dataset-specific tuning, with transfer to unseen datasets requiring further validation. We appreciate the reviewer's thoughtful feedback. Should any further questions or concerns arise, we would be happy to clarify them and respond promptly. **References for This Response** **[G1]** S. Segarra, A. G. Marques, and A. Ribeiro. Optimal Graph-Filter Design and Applications to Distributed Linear Network Operators. *IEEE Transactions on Signal Processing*, 65(15):4117-4131, 2017. **[G2]** F. Gama, B. G. Anderson, and S. Sojoudi. Node-Variant Graph Filters in Graph Neural Networks. *IEEE Data Science and Learning Workshop*, pp. 1-6, 2022. **[G3]** Y. Shin, J. Choi, S. Kim, and N. Park. TV-Rec: Time-Variant Convolutional Filter for Sequential Recommendation. *NeurIPS*, 2025. **[G4]** Z. Wu, S. Pan, G. Long, J. Jiang, and C. Zhang. Graph WaveNet for Deep Spatial-Temporal Graph Modeling. *IJCAI*, pp. 1907-1913, 2019. **[G5]** Z. Wu, S. Pan, G. Long, J. Jiang, X. Chang, and C. Zhang. Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks. *KDD*, pp. 753-763, 2020.
Reviewer_GMH2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TVTS, a multivariate time-series forecasting framework that models time-varying temporal patterns and directional dependencies among variables. It introduces a Time-Varying Convolutional Filter (TVCF) with a structured spectral basis for adaptive temporal filtering, and Directed Geometric Clustering (DGC) to construct sparse, asymmetric variable relationships for topology-guided attention. A frequency-domain residual prediction mechanism further refines the forecasts. Experiments across 17 datasets demonstrate competitive performance, supported by ablation studies, efficiency analyses, and directional relationship evaluations.

Critical strengths

Strength 1: A meaningful joint model of two forecasting challenges. TVTS handles changing temporal patterns and asymmetric relations within one framework. TVCF and DGC serve distinct purposes. This combination makes sense for applications such as traffic, energy, and weather forecasting.

Strength 2: Broad experiments and targeted ablations. Across many datasets, the results are strong. TVTS leads on 15 of the 16 aggregate long-term metrics in Table 2 and all eight aggregate PEMS metrics in Table 3. Table 4 compares key modules with alternatives such as static filtering and bidirectional topology. Appendix H.4 adds five-seed statistical tests. Importantly, the authors also note that the Traffic MSE improvement is not statistically significant. Reporting this exception makes the evaluation more credible.

Critical weaknesses

Weakness 1: The meaning and validation of directed dependencies remain limited. DGC produces an asymmetric matrix, but this alone does not show that its edges reflect real information flow. Appendix J.2 compares learned directions with Granger-based directions. That is useful evidence, but it does not establish physical or causal links. It is also unclear whether the graph changes with each input window. It would help to explain what these edges represent and to compare DGC with other directed graph learners at the same sparsity level.

Weakness 2: The added value of position-dependent spectral filtering needs stronger isolation. TVCF learns coefficients tied to temporal positions. This differs from filters that change with each input sequence, a limitation the authors note in Section 6. Although the ablations suggest that TVCF helps, they do not fully show whether the gain comes from the spectral design or simply from extra position-specific parameters. Comparisons with simpler positional gating or convolution under matched budgets would clarify this point. A clearer account of the cost of complex-valued operations would also help.

Questions for authors

Q1 : Appendix J.2 compares the learned directions with Granger-based relationships. However, it remains unclear whether these directions reflect meaningful predictive dependencies or simply asymmetric model weights. Could the authors clarify how the learned graph should be interpreted and whether it changes across input windows?

Q2 : Table 4 shows that TVCF performs better than the static filter baseline. However, this does not fully explain whether the improvement comes from the proposed spectral design or from having more position-specific parameters. Could the authors clarify this distinction or provide a comparison with a simpler position-aware filter under similar model sizes?

Q3 : Equation (7) sets unselected edges to zero, while Equation (8) uses a masked attention mechanism. Could the authors clarify how unselected edges are excluded from the softmax in the actual implementation? Are they assigned negative infinity before normalization, and how are self-connections handled?

3 comments

We thank the reviewer for the constructive feedback. We address each question through methodological clarification and additional experimental evidence.

Q1 and Weakness 1: Interpretation and Validation of Directed Dependencies

Graph interpretation and input adaptivity. DGC learns a dataset-level directed predictive prior from static variable embeddings (Eq. 6). Under row-wise aggregation (Eq. 8), AijA_{ij} measures the affinity of receiver ii to sender jj, and Top-kk selects dominant relations for attention modulation. Pairs outside Top-kk retain background attention rather than being explicitly masked. Since the embeddings are input-independent, the learned topology is shared across windows, while window-dependent queries and keys dynamically adjust attention weights to each input. DGC therefore captures persistent directional predictive relevance rather than verified physical or causal connections.

Predictive effectiveness under controlled comparisons. To evaluate DGC against alternative affinity learning mechanisms, we replace only its affinity construction with Graph WaveNet-style (GWNet) [G1] and MTGNN-style [G2] learners. All variants share the TVTS backbone, Top-kk budget, normalization, and fusion mechanism. Embedding dimensions are adjusted to maintain comparable graph-specific parameter counts (PgP_g). These experiments compare affinity learners within TVTS rather than complete forecasting models. Table R1 reports results averaged over H∈[96,192,336,720]H\in[96,192,336,720] for long-term datasets and H∈[12,24,48,96]H\in[12,24,48,96] for PEMS. DGC achieves the lowest averaged MSE and MAE across all 12 datasets, including ECL (MSE: 0.157 vs. 0.164/0.168 for GWNet/MTGNN). These results demonstrate the predictive effectiveness of DGC under comparable experimental settings.

Directional correspondence with Granger references. We further examine the alignment between learned affinities and predictive lead--lag relationships by comparing Aij−AjiA_{ij}-A_{ji} with Gj→i−Gi→jG_{j\rightarrow i}-G_{i\rightarrow j}, following the row-receiver/column-sender convention. Table R2 reports direction agreement (DA, %) and Spearman correlation (ρ\rho) across four lag orders (p∈[3,6,12,24]p\in[3,6,12,24]), using the same training-set variables. Ties count as nonagreement, and Random uses fixed random affinities. DGC achieves higher DA than all controls at every lag on Traffic, ECL, and PEMS07. On ECL, DA ranges from 59.2% to 68.3%, with ρ\rho ranging from 0.139 to 0.273. Weather shows mixed results, reflecting dataset-dependent directional correspondence. Overall, these findings provide evidence of alignment with predictive lead--lag structures.

Table R1. Forecasting results for alternative affinity learners within TVTS (L=96L=96; PgP_g: graph-specific parameters).

DatasetGWNet PgP_gGWNet MSEGWNet MAEMTGNN PgP_gMTGNN MSEMTGNN MAEDGC PgP_gDGC MSEDGC MAE
ETTh12240.4370.4292100.4430.4332240.4280.426
ETTh22240.3720.3952100.3740.3962240.3630.390
ETTm12240.3790.3912100.3810.3922240.3750.386
ETTm22240.2750.3192100.2740.3182240.2700.315
Weather6720.2510.2726400.2560.2756720.2400.267
Exchange2560.3390.3942720.3410.3952560.3340.390
Traffic275840.4350.281281280.4380.282275840.4280.277
ECL102720.1640.255101100.1680.257102720.1570.249
PEMS03114560.1230.228112200.1280.233114560.1200.226
PEMS0498240.1130.21696900.1340.22698240.1120.215
PEMS07282560.1160.207288000.1150.207282560.1030.203
PEMS0854400.1520.23353380.1610.24154400.1450.229

Table R2. Directional correspondence with Granger references (L=H=96L=H=96; four lag orders).

DatasetMethodp=3p=3 DAp=3p=3 ρ\rhop=6p=6 DAp=6p=6 ρ\rhop=12p=12 DAp=12p=12 ρ\rhop=24p=24 DAp=24p=24 ρ\rho
WeatherGWNet27.5-0.09326.7-0.18826.7-0.19829.2-0.156
MTGNN46.70.10450.50.12048.30.11949.20.107
Random47.5-0.13548.3-0.11142.5-0.14043.3-0.146
DGC55.80.01950.00.09947.50.15146.70.158
TrafficGWNet39.20.09943.30.12542.50.11544.20.150
MTGNN44.2-0.10243.3-0.11342.5-0.05252.50.013
Random45.80.05851.70.06249.20.01850.80.052
DGC52.00.12254.50.11153.00.14556.70.160
ECLGWNet39.20.14838.30.13336.70.10741.70.173
MTGNN45.8-0.07245.8-0.03950.00.01046.70.021
Random46.7-0.07248.3-0.06850.8-0.04154.2-0.066
DGC59.20.13965.80.18668.30.22465.00.273
PEMS07GWNet35.8-0.05637.5-0.06334.2-0.10138.3-0.084
MTGNN50.00.07652.50.07050.80.02950.00.032
Random54.2-0.09051.7-0.05848.3-0.06345.8-0.074
DGC60.00.18957.50.14655.80.11055.00.139

Due to space constraints in individual response boxes, we will address the remaining questions separately in the subsequent responses.

Q2 and Weakness 2: Position Dependence versus Additional Capacity

Controlled comparisons with position-aware alternatives. To determine whether TVCF's gains simply result from additional position-specific parameters, we compare three alternatives: (i) a position-shared nonlinear convolutional filter bank (Static Bank), (ii) the same bank with learnable positional gates (Gated Bank), and (iii) directly learned, unconstrained position-specific real-valued kernels (Direct Taps). All variants share the same backbone, temporal support, normalization, and training budget. Bank widths approximate TVCF's filter-parameter budget, while Direct Taps retains its native capacity. Since these alternatives lack spectral bases, basis orthogonality regularization is applied only to TVCF. These comparisons therefore evaluate complete filter parameterizations rather than spectral factorization in isolation.

Performance comparison. Table R3 reports results averaged over H∈[96,192,336,720]H\in[96,192,336,720] for long-term datasets and H∈[12,24,48,96]H\in[12,24,48,96] for PEMS. TVCF achieves the lowest MSE and MAE across all 12 datasets. For example, TVCF achieves MSEs of 0.157 on ECL and 0.428 on Traffic, compared with 0.161/0.162/0.161 and 0.441/0.445/0.443 for Static Bank/Gated Bank/Direct Taps, respectively. These results show that positional gating and unconstrained position-specific kernels alone do not match TVCF's predictive performance.

Parameter capacity and computational cost. Let T=LpadT=L_{\mathrm{pad}}. A TVCF filter contains M[T+2(K+1)]M[T+2(K+1)] real-valued parameters, accounting for its complex spectral bases, whereas Direct Taps contains L(K+1)L(K+1). For L=96L=96, K=48K=48, and M=64M=64, the respective counts are 15,424 and 4,704. A nonlinear bank with ww branches contains w(K+3)+1w(K+3)+1 parameters, with an additional LL parameters for positional gating. These counts refer to the filtering modules rather than the entire model. Thus, bank-based comparisons use approximately matched parameter budgets, whereas Direct Taps retains a smaller capacity. Moreover, parameter counts do not fully reflect the computational cost of complex-valued operations; Figure 6 and Table 23 provide complementary end-to-end efficiency measurements.

Table R3. Forecasting performance of temporal-filter alternatives within TVTS (L=96L=96).

DatasetStatic Bank MSEStatic Bank MAEGated Bank MSEGated Bank MAEDirect Taps MSEDirect Taps MAETVCF MSETVCF MAE
ETTh10.4340.4300.4470.4360.4350.4300.4280.426
ETTh20.3660.3940.3680.3950.3670.3930.3630.390
ETTm10.3790.3890.3800.3890.3780.3890.3750.386
ETTm20.2740.3190.2750.3180.2740.3180.2700.315
Weather0.2450.2680.2470.2700.2460.2690.2400.267
Exchange0.3360.3930.3380.3940.3370.3920.3340.390
Traffic0.4410.2810.4450.2830.4430.2830.4280.277
ECL0.1610.2540.1620.2540.1610.2530.1570.249
PEMS030.1240.2300.1260.2290.1260.2300.1200.226
PEMS040.1180.2180.1200.2200.1160.2190.1120.215
PEMS070.1080.2050.1100.2050.1060.2050.1030.203
PEMS080.1470.2320.1480.2330.1470.2330.1450.229

Q3: Unselected Edges and Self-Connections

Clarification of the implemented attention mechanism. In the actual implementation, unselected edges are assigned zero affinity scores rather than −∞-\infty, so they retain nonzero weights after row-wise softmax. Let Ak\mathbf{A}_k denote the Top-kk affinity matrix, with unselected entries set to zero. The implementation computes:

M=Softmax⁡row(Ak).\mathbf{M}=\operatorname{Softmax}_{\mathrm{row}}(\mathbf{A}_k).

=\operatorname{Softmax}_{\mathrm{row}}\left( \frac{QK^\top\odot\mathbf{M}}{\sqrt{d_k}} \right)V.$$ Accordingly, Top-$k$ determines affinity-based attention modulation rather than imposing a hard connectivity mask. Self-connections follow the same selection rule without mandatory inclusion; even when unselected, they remain accessible through nonzero attention weights and residual connections. We will revise Eq. (8) and its accompanying description to accurately reflect the implementation. Since the resulting attention remains dense, we make no claim of computational savings from sparse attention.

Weakness 2: Computational Cost of Complex-Valued Operations

Empirical computational cost. We benchmark warmed-up training steps on ECL-shaped inputs (b=16b=16, L=H=96L=H=96, N=321N=321). Each step includes forward propagation, loss computation, backpropagation, and an Adam update, excluding initialization, data transfer, and initial eigendecomposition. Complex parameters are counted as two real-valued scalars. Module-level benchmarks use independent local objectives; therefore, their runtimes and memory measurements are not additive. As reported in Table R4, TVCF requires 15.749 ms per step and 80.547 MiB peak memory, while the full TVTS model requires 23.605 ms and 502.650 MiB, respectively.

Implementation-level complexity. TVCF performs dense complex-valued GFT operations during training, while its cached real-valued T×TT\times T inference operator incurs O(bNT2)\mathcal{O}(bNT^2) operations per layer. FDRP involves rFFT/iRFFT and complex-valued linear transformations, while TGF computes dense N×NN\times N attention scores. We will refine the complexity analysis to reflect these actual implementation costs rather than relying solely on FFT-based estimates, and clarify that Top-kk selection does not reduce the asymptotic complexity of dense attention.

Table R4. Training cost of TVTS components on ECL-shaped inputs.

ModuleParamsTime (ms)Alloc. (MiB)Peak (MiB)
TVCF12416015.74925.88580.547
DGC + TGF431577613.16372.342402.777
FDRP10788312.09524.204128.814
Full TVTS464737323.60576.369502.650

We appreciate the reviewer's thoughtful feedback. Should any further questions or concerns arise, we would be happy to clarify them and respond promptly.

References for this response

[G1] Z. Wu, S. Pan, G. Long, J. Jiang, and C. Zhang. Graph WaveNet for Deep Spatial-Temporal Graph Modeling. IJCAI, pp. 1907-1913, 2019.

[G2] Z. Wu, S. Pan, G. Long, J. Jiang, X. Chang, and C. Zhang. Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks. KDD, pp. 753-763, 2020.

Comments on the paper

1 comment
Area_Chair_TemfOpenReview ↗

Discussion Reminder

Dear Reviewers,

Thank you for your efforts so far. This is a friendly reminder to review the authors' rebuttal and respond to their comments as soon as possible, before the discussion deadline (November 18 AoE). Please also indicate whether your concerns have been addressed. Thank you for your time and cooperation!

Best regards,
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.