acceptodds
Withdrawn from ICLR 2027

ChronoSheaf: Joint Spatio-Temporal Spectral Filtering on Cellular Sheaves

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sheaf neural networks generalize graph neural networks by assigning each edge a learned linear restriction map that specifies how the local representations of adjacent nodes relate to one another, thereby mitigating oversmoothing. Existing sheaf architectures, however, define the sheaf Laplacian as an instantaneous spatial operator and therefore cannot explicitly model how signals evolve over time on a graph. We first introduce a spectral sheaf Laplacian, defined jointly over space and time, such that a single operator regulates signal propagation across both the spatial graph and the temporal domain. Directly extending this operator's receptive field in space or time would require increasingly complex sheaf structures. Instead, we propose a spectral filter that expands the effective receptive field along both dimensions without enlarging the underlying sheaf. Building on this operator, we introduce ChronoSheaf, a spatio-temporal sheaf network with two key components. First, Transitional Restriction Maps (TRM) decompose each node's restriction map into a spatial term and two temporal frames that independently control its relation to neighboring nodes at the preceding and following timestep, enabling asymmetric temporal influence. Second, the Dual Spectral Filter (DSF) applies a single bounded-degree polynomial filter with an adaptively estimated spectral radius to the joint spatio-temporal sheaf operator. We validate ChronoSheaf on a diverse set of spatio-temporal forecasting benchmarks, where it outperforms state-of-the-art baselines. The code is available at: https://anonymous.4open.science/r/ChronoSheaf-1D7D/ .

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Cellular Sheaves, Graph Neural Networks, Graph Topology, Learning on Graphs

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_bgBHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper motivates ChronoSheaf by arguing that existing sheaf neural networks operate exclusively in the spatial domain and therefore cannot capture temporal dependencies through their sheaf Laplacians. It combines spatial and temporal relationships into a single sheaf Laplacian using restriction maps. It then applies a Chebyshev polynomial filter to capture information across multiple spatial hops and time steps. The authors claim this approach reduces oversmoothing and preserves node representations.

Critical strengths

Attempted to design a filtering operation to solve existing sheaf neural networks, which operate exclusively in the spatial domain and therefore cannot capture temporal dependencies through their sheaf Laplacians.

Critical weaknesses

1.The construction of the joint Laplacian in Equation (8) lacks sufficient mathematical and methodological justification. In particular, (Lin)(L_{\mathrm{in}}) and (Lout)(L_{\mathrm{out}}) cannot individually be considered valid symmetric positive-semidefinite Laplacians, as each contains only one direction of the temporal coupling.

  1. Simply adding spatial and temporal operators does not guarantee that the resulting spectral representation is optimal or even beneficial for the forecasting task. The authors provide neither a theoretical justification nor sufficient controlled experiments demonstrating why this construction is preferable to separately learned spatial and temporal operators or existing joint spatio-temporal filtering methods.

  2. The proposed dual spectral filter has a serious limitation. Recent papers, including HMH (ICML 2026) and Sign is Not a Remedy (ICML 2024), have discussed limitations of graph message passing, particularly frequency mixing and oversmoothing because of polynomial approximation. However, the authors simply add the spatial and temporal Laplacians and apply a single polynomial filter. Mathematically, when the operators share eigenvectors, the filter becomes g(λs+λt)g(\lambda_s+\lambda_t), meaning different spatial and temporal frequencies can receive the same response. This may mix important frequency information rather than preserve it. The authors neither discuss this limitation nor provide sufficient theoretical or experimental justification for their filtering design.

  3. The oversmoothing analysis in Section 4.1 is not convincing. The authors measure ordinary Dirichlet energy, although their proposed model operates in a learned sheaf space. A nontrivial sheaf can have zero sheaf energy while maintaining different node representations. Therefore, preserving ordinary Dirichlet energy does not establish that the model avoids oversmoothing in the sheaf sense.

  4. The proposed transitional restriction has been described as dynamically learned temporal relationships. But, Equation (4) only rescales a fixed node-specific matrix using a scalar gate. Therefore, the underlying transformation cannot change its orientation or row space across time. The same frame is also shared across neighboring edges, preventing edge-specific temporal transformations.

  5. The readability of the paper is poor, and the methodology is unnecessarily difficult to follow. The authors spend too much space presenting mathematical proofs instead of clearly explaining the proposed architecture, its intuition, and the motivation behind each component. Most of these proofs should be moved to the appendix. The main paper should focus on explaining how the method works, why each component is needed, and what it actually contributes to the learning task. The current presentation makes it difficult to understand and evaluate the main contributions.

  6. The authors use channel-dependent polynomial coefficients, but the learned sheaf Laplacian contains non-diagonal feature transformations. These operations generally do not commute, meaning the resulting filter is not necessarily diagonal in the Laplacian eigenbasis and may mix different spectral components. The authors should explain how this formulation preserves or controls spectral information and why it is preferable to conventional spectral filtering.

Questions for authors

The authors use channel-dependent polynomial coefficients, but the learned sheaf Laplacian contains non-diagonal feature transformations. These operations generally do not commute, meaning the resulting filter is not necessarily diagonal in the Laplacian eigenbasis and may mix different spectral components. Can you explain how this formulation preserves or controls spectral information and why it is preferable to conventional spectral filtering?

Reviewer_1fkoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors first introduce a spectral sheaf Laplacian defined jointly over space and time, enabling the modeling of spatial and temporal dependencies within a unified operator. Building on this formulation, they propose ChronoSheaf, a spatio-temporal sheaf network consisting of two main components: i) Transitional Restriction Maps (TRM), which decompose each node’s restriction map into a spatial component and two temporal frames, allowing asymmetric interactions with neighboring nodes at preceding and following timesteps; and ii) the Dual Spectral Filter (DSF), which applies a bounded-degree polynomial filter to the joint spatio-temporal sheaf Laplacian, using an adaptively estimated spectral radius to expand the receptive field across both dimensions. The authors establish theoretical properties of the proposed operator and evaluate ChronoSheaf on several real-world forecasting datasets spanning traffic, weather, and air quality forecasting.

Critical strengths
  • Rather than combining separate spatial and temporal modules, the proposed formulation introduces a unified spatio-temporal sheaf Laplacian that incorporates temporal interactions directly into the sheaf structure through distinct incoming and outgoing restriction maps. The authors demonstrate that the resulting operator preserves the symmetry and positive semi-definiteness properties characteristic of sheaf Laplacians, providing a mathematically grounded foundation for joint spatio-temporal spectral filtering
  • The proposed approach is evaluated on six real-world datasets spanning traffic, weather, and air quality forecasting, demonstrating strong predictive performance against several established and recent baselines. The evaluation covers different graph sizes and forecasting horizons, providing empirical evidence of the applicability of the proposed framework across diverse spatio-temporal settings
Critical weaknesses
  • Although the experimental results are promising, the evaluation does not fully isolate the benefits of the proposed sheaf formulation from those of joint spatio-temporal connectivity and polynomial filtering. While the ablation studies in Table 5 provide useful evidence for the contribution of individual architectural components, the extent to which the proposed sheaf formulation is responsible for the reported improvements remains insufficiently established
  • The parameterization of the Transitional Restriction Maps in Equations (3)–(4) limits their temporal variation to scalar rescaling of learned node-specific matrices. Consequently, within each layer, the incoming, outgoing, and spatial restriction maps retain their respective underlying linear transformations across timesteps. While this design may improve parameter efficiency, it restricts the expressiveness of individual time-dependent restriction maps. The paper does not investigate this trade-off or compare the proposed parameterization with more general feature-dependent restriction maps, making it difficult to assess whether the imposed restriction is beneficial or unnecessarily limiting
Questions for authors
  • Could the authors provide a comparison with a conventional spatio-temporal graph model using the same joint space-time graph, an equivalent Chebyshev filter, and comparable model capacity?
  • Could the authors clarify the motivation behind the scalar-gated parameterization of the restriction maps in Equations (3)–(4) and whether more expressive feature-dependent alternatives were considered?
  • Although previous work has shown that neural sheaf diffusion can mitigate oversmoothing compared to conventional GNNs [1], the proposed architecture also incorporates residual connections and layer normalization, which may independently influence the preservation of node representations [2]. Could the authors provide an ablation to clarify how much of the observed behavior can be attributed specifically to the sheaf structure rather than to these architectural components?
  • The authors report averaging results over five independent runs; however, the experimental tables only provide mean values, without standard deviations or confidence intervals. Could the authors report these uncertainty estimates to help assess the robustness of the reported improvements, particularly when performance differences between methods are small?

[1] Bodnar, Cristian, et al. "Neural sheaf diffusion: A topological perspective on heterophily and oversmoothing in gnns." Advances in Neural Information Processing Systems 35 (2022): 18527-18541.

[2] Scholkemper, Michael, et al. "Residual connections and normalization can provably prevent oversmoothing in gnns." International Conference on Learning Representations. Vol. 2025. 2025.

Reviewer_6pc4OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes ChronoSheaf, a spatio-temporal forecasting framework that extends cellular sheaves to a joint space-time graph. It introduces Transitional Restriction Maps to model heterogeneous interactions and a Dual Spectral Filter to capture higher-order spatio-temporal dependencies.

Critical strengths

(1) The paper addresses a relevant and important problem of modeling heterogeneous spatio-temporal dependencies in graph-based forecasting.

(2) The experiments provide useful empirical observations regarding forecasting performance, oversmoothing behavior, and spectral filter order.

Critical weaknesses

(1) Insufficient differentiation from existing joint spatio-temporal graph forecasting methods. Joint spatio-temporal modeling has been extensively investigated in previous work, including [1, 2]. In particular, [1] explicitly models cross-node, cross-time dependencies, while [2] explores spatial heterogeneity and dynamic temporal relationships. More recently, [3] has explored dynamic sheaf-based spatio-temporal forecasting. Although the paper acknowledges some earlier approaches, it does not adequately distinguish its fundamental contributions from these closely related frameworks.

(2) Limited methodological novelty in heterogeneous graph representation and spectral filtering. The methodological novelty is insufficiently differentiated from prior research on heterogeneous graph learning and adaptive spectral filtering, including [4, 5]. More importantly, [6] has already explored higher-order polynomial filtering on sheaf Laplacians, including Chebyshev recursions and spectral rescaling. The proposed method largely builds upon these established techniques.

(3) Limited theoretical justification for the claimed modeling advantages. The theoretical analysis mainly establishes symmetry, positive semi-definiteness, and receptive-field bounds, without sufficiently demonstrating improved expressive power or advantages over existing formulations.

(4) Insufficient empirical validation of the core design choices. Although the reported forecasting results are promising, the experiments do not sufficiently isolate the benefits of the proposed joint spatio-temporal sheaf formulation. First, the comparisons omit closely related joint spatio-temporal graph models and sheaf-based forecasting approaches. Second, the ablations in Table 5 are restricted to Molene and Air Quality and do not include the main traffic benchmarks or a controlled comparison against a decoupled spatial-sheaf and temporal-modeling architecture. Third, the oversmoothing analysis in Figure 4 does not isolate the contribution of sheaf propagation from other architectural factors, such as residual connections.

[1] Spatio-Temporal Joint Graph Convolutional Networks for Traffic Forecasting

[2] Learning Dynamics and Heterogeneity of Spatial-Temporal Graph Data for Traffic Forecasting

[3] Sheaf Diffusion with Adaptive Local Structure for Spatio-Temporal Forecasting

[4] Beyond Homophily and Homogeneity Assumption: Relation-Based Frequency Adaptive Graph Neural Networks

[5] Graph Neural Networks with Diverse Spectral Filtering

[6] Polynomial Neural Sheaf Diffusion: A Spectral Filtering Approach on Cellular Sheaves

Questions for authors

see weaknesses

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

Related papers

Discussion (0)

Sign in to comment.