- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper addresses cross-channel dependency modeling in multivariate time series forecasting and proposes SICDNet, which consists of two modules. (1) SICM replaces dense attention and adaptive graph-based methods with a dyadic, multi-stage/multi-round cascade of grouped Givens rotations, including cyclic shifts. It performs orthogonal, energy-preserving pairwise mixing along the variable dimension of inverted tokens. The rotation angles are dynamically generated by a small MLP from pairwise energies that are invariant to the rotations themselves. This enables exact inversion without caching intermediate angles. The computational complexity is reduced from to . (2) SGM applies a Type-II discrete cosine transform (DCT-II) along the latent dimension, combines a learnable static frequency response with data-dependent gating, and then performs multiplicative fusion between the frequency and feature domains. The experiments cover 13 datasets and 9 baselines. The authors also evaluate the generalizability of SICM by integrating it into DUET, iTransformer, SOFTS, and VCformer; conduct identity-control causal intervention analyses; visualize rotation angles; and investigate hyperparameter sensitivity.
Critical strengths
1: The design of SICM is interesting and technically well motivated. The main idea is to generate rotation angles from pairwise energies that remain unchanged under the rotations. This allows SICM to perform input-dependent orthogonal mixing while retaining exact invertibility without storing intermediate angles. The combination of dynamic conditioning and invertible transformations is a useful design choice for time-series channel modeling. The identity-control experiment also provides evidence that cross-channel mixing contributes to the results: setting the rotation angles to zero increases MAE, and the reported bootstrap confidence interval remains above zero. 2: The experiments provide useful evidence that SICM can work with different forecasting architectures. The authors integrate the module into DUET, iTransformer, SOFTS, and VCformer, rather than evaluating it only within their own model. Using the official hyperparameters and a consistent experimental setup makes these comparisons more informative. The paper also reports cases where the improvements are unstable, particularly on Solar and Traffic, instead of presenting the results as uniformly positive. This makes the evaluation more convincing.
Critical weaknesses
1: The relationship to existing methods needs to be clarified. SICM uses a dyadic cascade of Givens rotations, which is closely related to butterfly-style factorizations of structured linear transformations (e.g., Dao et al., ICML 2019). Its orthogonality and invertibility also connect it to established invertible neural networks. Meanwhile, SGM shares similarities with frequency-domain approaches such as FreTS, FNet, and GFNet. The paper does not discuss these connections, making it difficult to determine what is genuinely new. The main distinction appears to be the use of rotation-invariant pairwise energies to generate input-dependent rotation angles. The authors should explain more clearly how this differs from existing methods and what additional benefits it provides. 2: The practical value of invertibility is not demonstrated, and some performance gains are small. Although exact inversion is emphasized as a key property of SICM, the experiments do not show how it can be used in practice, for example, to reduce activation memory during backpropagation. Without such results, it is unclear how much this property contributes beyond the mathematical design. In addition, several reported improvements are small relative to the reported variation across runs. SICDNet also performs worse than iTransformer on Traffic in terms of MSE. These results suggest that the advantage is not consistent across datasets, so the claims about its overall superiority should be more carefully qualified.
Questions for authors
1: Table 3 refers to validation-set averages, whereas Table 10 appears to report test-set results. Could the authors clarify which evaluation split was used? In addition, how were ties handled when calculating the first-place counts in Table 8 2: Section 3.3 describes the DCT as operating along the temporal dimension and motivates it through the concentration of seasonal patterns and low-frequency trends. However, the implementation appears to apply the transform along the latent feature dimension. Could the authors clarify this discrepancy and explain the motivation for applying the DCT along that dimension? 3: SICM's dyadic cascade of Givens rotations appears closely related to butterfly factorizations, while its invertibility connects it to existing reversible neural networks. Could the authors clarify the main technical differences from these approaches? In particular, beyond generating rotation angles from rotation-invariant pairwise energies, what does SICM offer that existing structured transformations do not? 4: Section 3.3 describes the DCT as operating along the temporal dimension and motivates it through the concentration of seasonal patterns and low-frequency trends. However, the implementation appears to apply the transform along the latent feature dimension. Could the authors clarify this discrepancy and explain the motivation for applying the DCT along that dimension?