- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes COMPASS for multivariate time series forecasting. It builds on CycleNet’s learnable periodic embeddings and adjusts their contribution separately for each channel. The Channel-wise Relationship Calibration module compares the cross-channel correlation structures of the periodic prior and current input to generate calibration weights. Dynamic-guided Prior Enhancement then uses current features to aggregate periodic information across channels, followed by adaptive fusion and forecasting. Experiments cover eight long-term and four short-term forecasting datasets. The paper reports an average MSE reduction of 2.23% over the strongest competing methods, with additional ablations, comparisons across random seeds, and efficiency measurements.
Critical strengths
-
The paper addresses a reasonable limitation of periodic forecasting, as historical periodic patterns may be useful for some channels and less useful for others. The proposed architecture provides an explicit way to adjust their contribution and retrieve relevant periodic information from other channels. This is a useful extension of a learned periodic-prior model.
-
The evaluation covers 12 datasets and several forecasting horizons, including comparisons with recent periodic forecasting methods. The improvement on PEMS08 is substantial, and the experiments across four random seeds on three datasets provide additional support for the forecasting results. Reporting parameter counts and training time also helps assess the cost of introducing cross-channel modeling.
Critical weaknesses
-
The claimed decoupling effect of CRC is not explained by the operation as written. Equation (6) multiplies each channel of the prior and input by positive scalar weights. For a fixed input window, Pearson correlation is invariant to such scaling: $ \operatorname{corr}(s_i\pi_i,(1-s_i)x_i) =\operatorname{corr}(\pi_i,x_i). $ The same applies to cross-channel correlations under positive channel-wise scaling. Figures 6 and 8 nevertheless show reduced Pearson correlation after CRC and interpret this as evidence of reduced redundancy. If the measurements include the learned projections, subsequent FFNs, or pooling across windows with different weights, the correlations may change, but the effect cannot then be attributed to the scalar calibration alone. This matters because decoupling the prior and current dynamics is a central explanation of the method. Appendix D also assumes that larger relationship mismatch indicates greater prior uncertainty, which does not establish or empirically validate this connection.
-
The evidence for the specific calibration design is weaker than the overall forecasting results suggest. In Table 3, replacing channel-wise scores with a shared score gives nearly identical results on ETTh1 and identical MSE at the reported precision on Solar and PEMS03. Cosine-based calibration also matches the full model on several datasets. These comparisons need uncertainty estimates to establish the benefit of the proposed correlation-discrepancy measure. More importantly, Table 7 labels its final dataset as PEMS08 but reports COMPASS MSE 0.086 and MAE 0.185, matching PEMS03 rather than the PEMS08 results in Tables 2 and 6. This leaves the ablation evidence for the dataset with the largest improvement unclear. The reported gains are also uneven. For example, PEMS08 improves substantially, while Traffic MSE is 0.455 compared with 0.409 for SOFTS.
Questions for authors
-
Which tensors and axes are used to compute the correlations in Figures 6 and 8? Are these correlations computed within individual windows or after pooling multiple windows? Please explain how the reported reduction arises given the invariance of Pearson correlation to positive scalar weighting in Eq. (6).
-
What evidence shows that the relationship discrepancy measures the usefulness of the periodic prior for a particular channel? For example, does a larger predict a larger forecasting error when relying on that prior? The uncertainty relationship in Appendix D is currently an assumption.
-
Does the final block of Table 7 correspond to PEMS03 or PEMS08? Please provide the correct PEMS08 ablations and results across multiple seeds for the full model, shared-score calibration, and cosine-based calibration.
-
Please clarify the implementation and training settings. In particular, Eq. (9) uses values derived from the refined prior , whereas Algorithm 1 uses . Is constrained to remain positive as required by Appendix D, and how were the model configurations and baseline hyperparameters selected?
Thank you for the authors' response, which has addressed most of my concerns. While I have already assigned a positive score, the manuscript still contains several minor issues that require further revision. Therefore, I will maintain my original score.