- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes DCNet, a dual-context framework for multivariate time-series forecasting. It combines global periodic patterns with local historical observations through a modified attention mechanism, using mixed queries and an expanded key–value space. A SwiGLU-based feed-forward module is also incorporated. Experiments across 12 datasets show competitive forecasting performance, with additional ablations, efficiency comparisons, and noise robustness evaluations. The results are encouraging, especially on high-dimensional PEMS datasets, although I have some concerns about the methodological novelty.
Critical strengths
-
The paper addresses an important problem in multivariate forecasting. Combining global periodic information with local observations is a reasonable approach to modeling variable dependencies.
-
I find the empirical results encouraging. DCNet achieves top-two performance on 22 of 24 aggregated metrics, with particularly strong results on the PEMS datasets.
-
The ablation experiments support the usefulness of the dual-context mechanism. I also appreciate the additional efficiency and noise robustness analyses.
Critical weaknesses
-
Novelty is somewhat limited. TQNet already uses global periodic representations to guide attention, while CycleNet provides the underlying periodic modeling approach. DCNet mainly extends these ideas through query mixing and key–value concatenation.
-
The claim about reducing spurious correlations is not fully supported. Clearer attention maps do not necessarily indicate more meaningful variable relationships.
-
The ablation study could be more specific. It is unclear how much improvement comes from query mixing versus simply adding periodic tokens to the attention mechanism.
-
The robustness evaluation is relatively narrow. Noise experiments are conducted only on Electricity, and the comparison does not include TQNet.
Questions for authors
-
What is the main methodological difference between DCNet and TQNet? Could the authors provide a more direct comparison under matched model capacity and experimental settings?
-
Could the authors separately evaluate query mixing and key–value expansion? I would like to understand which component contributes most to the improvements.
-
How can the authors demonstrate that DCNet actually reduces spurious inter-variable correlations rather than simply changing attention patterns?
-
Would the robustness improvements remain similar on other datasets and against stronger baselines such as TQNet? How sensitive is DCNet to incorrectly specified periodic lengths?
5 comments
The complete results are reported below. Each cell is MSE/MAE.
| Dataset | Step | DCNet | Q | KV | MLP | Q+MLP | KV+MLP | Q+KV | W/o Both | TQNet | iTransformer |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Metric | - | MSE/MAE | MSE/MAE | MSE/MAE | MSE/MAE | MSE/MAE | MSE/MAE | MSE/MAE | MSE/MAE | MSE/MAE | MSE/MAE |
| Electricity | 96 | 0.131/0.221 | 0.134/0.229 | 0.138/0.233 | 0.145/0.237 | 0.133/0.227 | 0.136/0.228 | 0.134/0.229 | 0.145/0.238 | 0.134/0.229 | 0.148/0.240 |
| - | 192 | 0.150/0.238 | 0.155/0.247 | 0.156/0.247 | 0.163/0.254 | 0.153/0.243 | 0.153/0.242 | 0.152/0.245 | 0.163/0.256 | 0.154/0.247 | 0.162/0.253 |
| - | 336 | 0.165/0.254 | 0.168/0.265 | 0.171/0.264 | 0.177/0.269 | 0.168/0.266 | 0.166/0.261 | 0.168/0.265 | 0.178/0.273 | 0.169/0.264 | 0.178/0.269 |
| - | 720 | 0.192/0.279 | 0.208/0.301 | 0.204/0.297 | 0.207/0.298 | 0.199/0.289 | 0.194/0.283 | 0.221/0.311 | 0.228/0.312 | 0.201/0.294 | 0.225/0.317 |
| - | Avg | 0.159/0.248 | 0.166/0.261 | 0.167/0.261 | 0.173/0.265 | 0.163/0.256 | 0.162/0.254 | 0.169/0.263 | 0.179/0.270 | 0.165/0.259 | 0.178/0.270 |
| PEMS03 | 12 | 0.056/0.155 | 0.058/0.159 | 0.061/0.159 | 0.064/0.166 | 0.059/0.160 | 0.059/0.159 | 0.058/0.159 | 0.064/0.167 | 0.060/0.161 | 0.071/0.174 |
| - | 24 | 0.068/0.171 | 0.075/0.180 | 0.074/0.178 | 0.081/0.188 | 0.075/0.181 | 0.072/0.176 | 0.069/0.174 | 0.082/0.190 | 0.077/0.182 | 0.093/0.201 |
| - | 48 | 0.089/0.196 | 0.102/0.211 | 0.098/0.204 | 0.114/0.225 | 0.098/0.206 | 0.098/0.203 | 0.091/0.198 | 0.114/0.225 | 0.104/0.215 | 0.125/0.236 |
| - | 96 | 0.126/0.229 | 0.144/0.247 | 0.136/0.236 | 0.154/0.265 | 0.145/0.246 | 0.133/0.234 | 0.136/0.235 | 0.162/0.276 | 0.148/0.253 | 0.164/0.275 |
| - | Avg | 0.084/0.187 | 0.095/0.199 | 0.092/0.194 | 0.103/0.211 | 0.094/0.198 | 0.091/0.193 | 0.089/0.192 | 0.106/0.215 | 0.097/0.203 | 0.113/0.222 |
| PEMS04 | 12 | 0.063/0.161 | 0.065/0.164 | 0.067/0.166 | 0.073/0.176 | 0.064/0.162 | 0.065/0.163 | 0.065/0.165 | 0.074/0.176 | 0.067/0.166 | 0.078/0.183 |
| - | 24 | 0.072/0.171 | 0.075/0.177 | 0.078/0.178 | 0.089/0.195 | 0.075/0.175 | 0.079/0.179 | 0.074/0.175 | 0.089/0.195 | 0.077/0.181 | 0.095/0.205 |
| - | 48 | 0.083/0.185 | 0.096/0.206 | 0.091/0.195 | 0.115/0.223 | 0.091/0.196 | 0.089/0.192 | 0.086/0.196 | 0.117/0.227 | 0.097/0.206 | 0.120/0.233 |
| - | 96 | 0.097/0.201 | 0.126/0.239 | 0.107/0.214 | 0.145/0.252 | 0.123/0.233 | 0.101/0.208 | 0.104/0.209 | 0.147/0.256 | 0.123/0.233 | 0.150/0.262 |
| - | Avg | 0.078/0.179 | 0.091/0.197 | 0.086/0.188 | 0.106/0.212 | 0.088/0.192 | 0.084/0.186 | 0.082/0.186 | 0.107/0.214 | 0.091/0.197 | 0.111/0.221 |
| PEMS07 | 12 | 0.050/0.140 | 0.052/0.143 | 0.053/0.145 | 0.056/0.148 | 0.052/0.142 | 0.051/0.141 | 0.052/0.143 | 0.057/0.151 | 0.051/0.143 | 0.067/0.165 |
| - | 24 | 0.058/0.150 | 0.064/0.161 | 0.064/0.158 | 0.069/0.166 | 0.065/0.163 | 0.061/0.154 | 0.059/0.152 | 0.073/0.173 | 0.063/0.159 | 0.088/0.190 |
| - | 48 | 0.070/0.163 | 0.080/0.178 | 0.078/0.171 | 0.091/0.189 | 0.077/0.173 | 0.077/0.171 | 0.073/0.168 | 0.097/0.201 | 0.081/0.179 | 0.110/0.215 |
| - | 96 | 0.087/0.180 | 0.099/0.197 | 0.093/0.188 | 0.113/0.204 | 0.096/0.192 | 0.094/0.189 | 0.091/0.187 | 0.123/0.227 | 0.103/0.203 | 0.215/0.245 |
| - | Avg | 0.066/0.158 | 0.074/0.170 | 0.072/0.166 | 0.082/0.177 | 0.073/0.168 | 0.071/0.164 | 0.069/0.163 | 0.088/0.188 | 0.075/0.171 | 0.120/0.204 |
| PEMS08 | 12 | 0.067/0.162 | 0.073/0.172 | 0.068/0.164 | 0.077/0.171 | 0.071/0.170 | 0.068/0.163 | 0.069/0.165 | 0.078/0.174 | 0.071/0.170 | 0.079/0.182 |
| - | 24 | 0.083/0.174 | 0.097/0.192 | 0.086/0.180 | 0.105/0.197 | 0.093/0.187 | 0.086/0.178 | 0.083/0.174 | 0.109/0.202 | 0.096/0.196 | 0.115/0.219 |
| - | 48 | 0.109/0.191 | 0.138/0.224 | 0.117/0.200 | 0.153/0.231 | 0.139/0.226 | 0.115/0.198 | 0.115/0.198 | 0.154/0.238 | 0.149/0.244 | 0.186/0.235 |
| - | 96 | 0.138/0.205 | 0.179/0.252 | 0.149/0.219 | 0.181/0.256 | 0.166/0.237 | 0.146/0.214 | 0.147/0.216 | 0.188/0.263 | 0.253/0.309 | 0.221/0.267 |
| - | Avg | 0.099/0.183 | 0.122/0.210 | 0.105/0.191 | 0.129/0.214 | 0.117/0.205 | 0.104/0.188 | 0.104/0.188 | 0.132/0.219 | 0.142/0.230 | 0.150/0.226 |
The KV-only variant provides the largest single-component gain on all four PEMS datasets. On Electricity, the Q-only variant is slightly stronger than KV-only. The complete DCNet remains best on every dataset, indicating that the components are complementary rather than redundant.
Q3.How can the authors demonstrate that DCNet actually reduces spurious inter-variable correlations rather than simply changing attention patterns?
We thank the reviewer for this important correction. We agree that a clearer attention map does not, by itself, demonstrate that the learned variable relations are more meaningful. To our knowledge, none of the twelve benchmark datasets provides a verified variable-level causal graph. We therefore use a controlled synthetic benchmark with a known dependency graph.
The latent variables follow , where is a known sparse graph and is a periodic component with period . The observed input is : the disturbance changes the observed Pearson correlations across regimes but does not affect the future latent target. Thus, spurious correlations can be present in the input without being valid predictive edges.
We use 7 variables, 9 nonzero off-diagonal edges among 42 directed candidate pairs, look-back length 96, and prediction length 96. The experiment is repeated with three seeds. We construct a dependency matrix by perturbing one input variable at a time and measuring its first-step effect on each output variable. The resulting matrix is compared with the known graph. We report AUROC, the area under the receiver operating characteristic curve: it is the probability that a true edge receives a higher dependency score than a non-edge. Random ranking gives approximately 0.5; higher is better. We also report AUPRC, the area under the precision-recall curve, which focuses on true edges among the highest-ranked pairs. Since 9 of 42 candidate pairs are true, random ranking gives an AUPRC baseline of approximately .
| Model | Test MSE | Test MAE | Dependency AUROC | Dependency AUPRC |
|---|---|---|---|---|
| DCNet | 0.3899 +/- 0.0105 | 0.4934 +/- 0.0068 | 0.6790 +/- 0.0732 | 0.3621 +/- 0.0852 |
| iTransformer | 0.3990 +/- 0.0092 | 0.4992 +/- 0.0058 | 0.5039 +/- 0.0166 | 0.2888 +/- 0.0302 |
| TQNet | 0.4150 +/- 0.0086 | 0.5088 +/- 0.0055 | 0.5488 +/- 0.0101 | 0.2544 +/- 0.0070 |
DCNet obtains lower test MSE and MAE than both baselines. Its dependency AUROC is higher than iTransformer by 0.175 and than TQNet by 0.130. Its AUPRC is also higher than iTransformer by 0.073 and than TQNet by 0.108.
As an additional control, we evaluate both models with the correct phase and a shuffled phase.
| Model | Phase | AUROC | AUPRC | MSE degradation |
|---|---|---|---|---|
| DCNet | Correct | 0.6790 +/- 0.0732 | 0.3621 +/- 0.0852 | - |
| DCNet | Shuffled | 0.5533 +/- 0.0457 | 0.3176 +/- 0.0313 | 32.31% +/- 8.80% |
| TQNet | Correct | 0.5488 +/- 0.0101 | 0.2544 +/- 0.0070 | - |
| TQNet | Shuffled | 0.5488 +/- 0.0101 | 0.2544 +/- 0.0070 | 0.00% |
The TQNet temporal query and query/key projections collapse to nearly zero during training under this setting, making the learned query phase-insensitive.This also explains why TQNet, which incorporates periodic representation, has higher forecasting error than iTransformer. In contrast, phase corruption substantially degrades both DCNet's forecasting error and dependency recovery. These results indicate that DCNet effectively uses the phase-aligned periodic context and recovers the known dependency graph more closely than both a model without periodic context and the TQNet baseline.
Response to Reviewer gt3w
We thank the reviewer for the careful evaluation. To address your concerns, we will answer your questions one by one below.
Q1. What is the main methodological difference between DCNet and TQNet? Could the authors provide a more direct comparison under matched model capacity and experimental settings?
We appreciate TQNet as an effective and efficient prior method. Its two core components are (i) a multi-head attention module that uses a periodic query, operating over the look-back dimension L, and (ii) an MLP with two linear layers and two activations, operating over the hidden dimension D. This design is simple and effective, but its two modules use different feature axes. As a result, it is less naturally organized as a conventional stack of variable-token Transformer blocks.
DCNet instead adopts a tokenized inverted-Transformer structure and operates over the variable dimension N with hidden dimension D. Both tokenized local and periodic contexts participate in query construction and key/value retrieval:
Q=βXenc+(1−β)Penc,K,V=[Xenc;Penc].This allows DCNet to use standard variable-token encoder stacking when necessary. In several settings, a single encoder layer is already sufficient, but the standard stacking interface also makes the model practical for larger-scale datasets.
We also note a difference in portability. TQNet's transfer experiments apply the attention block itself, for example through additive fusion in PatchTST and DLinear. In contrast, the DC construction combines local and periodic contexts at the representation level, so it can also be applied to purely linear backbones. For example, on PEMS08, the DC-augmented CycleNet outperforms TQNet in Table 3 and Table 6. The DLinear variant improved with DC technology also significantly outperforms the corresponding TQNet transfer variant. These comparisons reflect different design goals: TQNet is intentionally compact, while DC is designed as a portable dual-context interface.
For quantitative comparisons, all models use the same data split, look-back length, prediction lengths, and training protocol. Baseline-specific settings follow their official implementations. Exact parameter matching is difficult because the architectures operate on different axes; we therefore report parameter counts and use within-DCNet component ablations to isolate the mechanism. The Electricity result in Table 2 and Table 8 uses a single encoder layer, which is comparable to TQNet's compact setting.
Q2. Could the authors separately evaluate query mixing and key–value expansion? I would like to understand which component contributes most to the improvements.
We added a component-level ablation over the composite query Q, the expanded key/value branch KV, and the MLP. The variants are: