acceptodds
Under review as a conference paper at ICLR 2027

Decompose Dynamics Before Learning Dependencies in Spatiotemporal Systems

Abstract

Relations in networked spatiotemporal systems are often learned from observations that entangle dynamics governed by different mechanisms, obscuring what evolves locally and how it propagates across nodes. We introduce Component-Aware Network Dynamics with Ordered Relations (CANDOR), which decomposes local dynamics before learning their dependencies. CANDOR represents each trajectory through a persistent background, gradual accumulation and release, and sparse shocks. Conditioned on these components, a delay-aware physical branch models edge and sample-dependent propagation over directed topology, while a topology-unconstrained functional branch discovers latent dependencies from background dynamics. Context-adaptive fusion combines functional, forward-propagation, and reverse-support forecasts, with training objectives encouraging specialized and semantically consistent representations. Experiments on two traffic benchmarks and three long-horizon water-quality datasets span two distinct spatiotemporal systems: human-driven urban traffic and naturally evolving river water quality. CANDOR consistently outperforms the strongest baselines, reducing MAE by up to 4.31% for traffic and MSE by up to 5.47% for water-quality forecasting. These results establish decomposition before dependency learning as an effective principle for spatiotemporal representation learning.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatiotemporal dynamics, structured decomposition, delay-aware propagation, adaptive graph learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_EtYXOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper introduces CANDOR for forecasting in networked dynamical systems. It decomposes historical representations into background, accumulation, release, and shock components. These representations support propagation over physical topology and functional dependency learning beyond that topology. Three forecast branches are combined through adaptive fusion. Experiments on two traffic benchmarks and three water quality datasets report consistent improvements over 15 baselines, supported by repeated runs, ablations, and additional analyses.

Critical strengths
  1. The empirical evaluation provides meaningful evidence of forecasting effectiveness. The paper compares against 15 baselines, reports results over three seeds, and evaluates two distinct application domains. The reported improvements reach 4.31% in traffic MAE and 5.47% in water quality MSE. The complete results and ablation studies in the appendix strengthen confidence that the gains extend beyond a single dataset or experimental setting. This is the strongest consideration supporting acceptance.
  2. The architecture offers a useful way to combine physical topology with dependencies beyond direct connectivity. The ablations show that the contributions of the physical and functional branches vary across datasets, which supports their complementary use. The model also remains compact, with fewer than 0.91 million parameters. These results give the approach practical value for forecasting systems with different interaction patterns.
Critical weaknesses
  1. The relationship to D2STGNN is insufficiently documented, which creates a substantial concern about attribution and originality. In the supplied implementation excerpts, CANDOR’s LocalizedSpatiotemporalConvolution closely follows D2STGNN’s STLocalizedConv, including temporal window extraction, graph power construction, local mean pooling, and graph convolution. CANDOR’s RelationCompatibility also closely follows DistanceFunction in its feature extraction pipeline and query and key computations. However, D2STGNN is absent from the manuscript’s citations and baseline comparisons. D2STGNN already combines residual decomposition, separate dynamics modeling, and dynamic graph learning, so these connections are directly relevant to the paper’s central contribution. The authors need to clarify the extent of reuse, the substantive modifications, and the newly introduced methods. The physical branch ablation establishes that the branch is useful, while leaving its contribution beyond the existing design unresolved.

  2. The rationale and specific effectiveness of the proposed decomposition remain insufficiently established. The paper does not clearly explain which system dynamics or statistical assumptions justify the four components and the use of current and subsequent background references. Learned extractors and auxiliary objectives specify the computation, but further evidence is needed to justify this particular partition. The existing ablations support the usefulness of the complete decomposition module, while leaving its advantage over similarly sized alternatives unresolved. The subsequent combination of the components also makes it difficult to distinguish gains from improved node features from gains in dependency learning. This matters because the paper presents decomposition before dependency learning as its central methodological principle.

Questions for authors
  1. Clarification of implementation and attribution:

Please specify which modules were reused or adapted from D2STGNN, particularly LocalizedSpatiotemporalConvolution and RelationCompatibility. Please clarify their provenance, identify the substantive modifications, and explain how the manuscript distinguishes existing components from new contributions.

  1. Targeted evidence for the claimed contribution:

Is there a comparison between CANDOR and D2STGNN using the same data splits, forecasting inputs, and evaluation protocol? Please also identify which controlled comparisons retain the same physical backbone and separately evaluate the contributions of the new decomposition and functional relationship modeling. This evidence would materially affect my assessment of the contribution beyond prior work.

  1. Clarification and targeted evidence for the decomposition:

Please explain the generative or statistical assumptions that motivate the four components and the shifted background reference. Are there comparisons with plausible alternative decompositions under similar capacity and the same forecasting backbone? Such evidence would help establish whether the reported gains depend on this specific partition and how it supports subsequent dependency learning.

5 comments
AuthorsOpenReview ↗

Weakness1: Relationship to D²STGNN and methodological contributions

Weakness1:

We address four aspects of this concern: the implementation relationship between CANDOR and D²STGNN, the omission of D²STGNN from the original citations and baseline comparisons, the substantive differences between their physical modules, and CANDOR's contributions beyond its physical branch.

1. Code reuse, adaptations, and attribution. The implementation of CANDOR's physical branch adapts parts of D²STGNN's foundational implementation. The original implementation retained D2-style comments identifying modules that referenced D²STGNN. The two methods differ substantially in their detailed implementation and overall computational architecture. To clarify this relationship, the supplementary material provides a class-by-class implementation comparison between CANDOR's model.py and D²STGNN, including the LocalizedSpatiotemporalConvolution and RelationCompatibility modules, together with architectural diagrams to show their differences in implementation and design. Our original baseline selection prioritized representative work from the preceding three years. D²STGNN fell outside that selection window and was consequently omitted from the original baseline set.

The revised manuscript now formally cites and discusses D²STGNN, includes it as a full-model baseline, and adds an ablation replacing CANDOR's physical module with the corresponding D²STGNN module. The README accompanying the code also acknowledges D²STGNN and the other implementation references. Through manuscript citations, experimental comparisons, code attribution, and the implementation comparison, we clarify the connection and specific differences between CANDOR and D²STGNN.

2. Physical-branch differences. The distinction concerns the propagated states, aggregation weights, and forecasting feedback:

  • Direction-specific states. D²STGNN applies forward, backward, and adaptive graph supports to a shared diffusion representation. CANDOR constructs separate forward and reverse states before aggregation, using direction-specific node-calendar gates, temporal encoders, and residuals. Each path maintains its own temporal memory and propagation residuals, making state construction and evolution direction-specific.

  • Shock modulation and offset-specific weights. CANDOR uses Shock to modulate the propagation before graph aggregation. D²STGNN adapts its graph to the input sample but shares a spatial support across positions within the local window. CANDOR uses

    Wd,τ=RowNorm⁡(Td⊙ωd,τ⊙ad),\mathbf W^{d,\tau}=\operatorname{RowNorm}(\mathbf T^d\odot\boldsymbol\omega^{d,\tau}\odot\mathbf a^d),

    where Td\mathbf T^d restricts physical edges, ωd,τ\boldsymbol\omega^{d,\tau} assigns edge affinities by temporal offset, and ad\mathbf a^d adapts node relations to the observed history. The propagation weights vary across both offsets and samples.

  • Stepwise recombination and shared feedback. D²STGNN rolls out diffusion and inherent forecasts separately before combining them. At each step, CANDOR combines the local temporal state and graph contribution in each direction:

    h^i,p+1d=Li,pd+oi,pd.\widehat{\mathbf h}_{i,p+1}^{d}=\mathbf L_{i,p}^{d}+\mathbf o_{i,p}^{d}.

    The two directional predictions are averaged and fed back to both paths. Each path then updates its own temporal memory and propagation residuals. This coordinates the forecasts while retaining separate directional state evolution.

  • Role supervision and functional relations. Role-consistency losses align next-patch Background with local temporal states, current Release with propagation residuals, and next-patch Accumulation with graph contributions. An independent functional branch uses Background to guide sparse multi-relation discovery beyond physical edges. Context-adaptive fusion combines the functional, forward-propagation, and reverse-support forecasts.

The supplementary material documents these differences through the implementation comparison and architectural diagrams.

3. Contributions beyond the physical branch. The physical branch is important, especially on traffic datasets. Our objective also includes a broadly applicable model that performs well across spatiotemporal domains. This is our primary motivation. Ablations show that removing physical propagation has a relatively small effect on BASIN030501 and BASIN031300, where the functional branch and four-role decomposition contribute more. This reflects differences in dynamics and predictive dependencies across systems. CANDOR combines complementary mechanisms without requiring one branch to dominate every setting, achieving the lowest mean primary error on all five datasets among the compared methods.

AuthorsOpenReview ↗

Weakness2: Dynamical motivation, capacity-matched comparisons, and dependency learning

Weakness2:

1. Dynamical motivation for the four roles. Appendix A.1 of the revised manuscript develops the construction from traffic conservation and river transport–reaction balances. Under a nonnegative transport allocation, a node's state has complementary receiving and releasing descriptions. The receiving description combines the locally retained contribution with incoming Accumulation. The releasing description combines the part retained for the subsequent state with outgoing Release. An additional forcing response contributes to both descriptions.

These two descriptions motivate Background, Accumulation, Release, and Shock, together with the current and subsequent Background references. Their difference links Accumulation and Release to background evolution. CANDOR uses this structure as an inductive bias for latent role extraction, with both references lying within the observed history. Appendix A.1 provides the assumptions and derivation connecting the physical balances to these reference-dependent roles.

2. Comparison with similarly sized alternative decompositions. We added a parameter-matched w/o Decomp. control and Moving Average, STL, TimeMixer-style, and Trend-Season (DLinear-style) alternatives across all five datasets. Trainable parameter counts differ from CANDOR by less than 0.08%. The shared prediction loss, branch prediction supervision, and historical prediction consistency are retained, while constraints tied to particular components or state-separation mechanisms follow the corresponding construction. Each added control uses three seeds.

CANDOR achieves the lowest mean primary error on all five datasets. Relative to the strongest alternative decomposition, the reductions are 1.37% on METR-LA, 2.97% on PEMS-BAY, 2.80% on BASIN030501, 0.90% on BASIN031300, and 0.63% on BASIN180102. The parameter-matched w/o Decomp. also remains worse across all five datasets. Detailed configurations, parameter counts, and complete results are provided in Appendix A.11.

3. How decomposition contributes to node representations and dependency learning. The decomposition organizes node features into dynamical roles, and the architecture uses those roles to guide relation construction and propagation. Recombining the components produces task-adaptive historical representations for forecasting and information exchange. Alongside this integration, role-specific pathways continue to guide the relational computations.

Background is retained separately to construct the Functional Branch's cross-node relations. Shock is retained separately to modulate the Physical Branch's propagation features before spatial aggregation. Role-consistency supervision further connects next-patch Background to local temporal states, current Release to propagation residuals, and next-patch Accumulation to incoming graph contributions. Therefore, the decomposition shapes both the representations exchanged between nodes and the learning of their interactions.

This connection explains how recombination fits the method's central principle. The contribution of decomposition is expressed through its use in the downstream architecture. Node representation learning, relation construction, and propagation are trained jointly, and their ablation effects describe contributions within this coupled framework. Appendix A.10 evaluates the physical, functional, and relation-learning mechanisms, while Appendix A.11 examines whether comparable capacity or conventional temporal components reproduce the forecasting benefits.

These comparisons support decomposition before dependency learning as a role-conditioned organization of local dynamics and cross-node interactions. The architecture specifies how the decomposed information is used, and the decomposition supplies the dynamical roles that guide this processing across different spatiotemporal systems. In this setting, CANDOR's forecasting advantage comes from the combined effects of the decomposition mechanism and the forecasting architecture.

AuthorsOpenReview ↗

Questions1-3: Implementation, controlled comparisons, and decomposition

Question1:

1. Implementation and attribution. We have responsed code reuse, source attribution, and CANDOR's contributions in our response to Weakness1.

Question2:

We have added a full-model comparison and a physical-module replacement experiment. Below, we specify what each ablation retains.

1. Full-model comparison across systems. We compare CANDOR and D²STGNN under the same data splits, forecasting inputs, and evaluation protocol (Appendix A.8).

We use the same CANDOR architecture across road traffic and river water-quality networks. CANDOR achieves improvements on both traffic and river datasets. These results support its cross-system adaptability that one model architecture performs effectively in two distinct spatiotemporal systems.

The detailed results are provided in Section 4.1 and Appendix A.9 of the revised manuscript.

2. Replacement of the entire physical module. We replace CANDOR's entire physical module with the D²STGNN-based module while retaining the four-role decomposition and functional branch. The replacement configuration is described in Section 4.2.

In paired comparisons over three seeds, CANDOR's original physical module achieves lower primary error in four of the five settings. Replacing it with the D²STGNN module:

  • Increases METR-LA MAE by approximately 1.0%.
  • Increases MSE on the three river datasets by approximately 0.3%-6.6% at the 192-hour horizon, with the largest increase on BASIN180102.
  • Decreases PEMS-BAY MAE by approximately 0.1%.

These comparisons support the effectiveness of CANDOR's integrated physical-module design across systems, with the clearest benefit on BASIN180102. The results are shown in Figure 8 of the revised manuscript.

3. What each ablation retains in the physical backbone. The configurations are defined as follows.

(1) w/o Decomp. This configuration retains directed, delay-aware graph construction and the bidirectional shared-feedback autoregressive rollout. It replaces the four-role representations with undecomposed patch features and removes decomposition-dependent processing, including Shock modulation.

(2) w/o Aux. loss. This configuration retains the decomposition module and the complete forward computation of the physical branch. It removes decomposition-specific auxiliary supervision, including dual-reference reconstruction and role alignment.

(3) w/o Functional, w/o Relation, and avg Fusion. Both w/o Functional and w/o Relation retain the physical module:

  • w/o Functional removes the entire functional branch.
  • w/o Relation disables cross-node functional relation learning while retaining the rest of that branch.
  • avg Fusion evaluates an alternative fusion configuration.

The results are reported in Figure 8 of the revised manuscript.

(4) w/o Physical. This configuration removes the entire physical backbone, leaving the functional branch to produce the forecast.

(5) Forward only. This configuration retains the forward physical path and removes the reverse path.

(6) Reverse only. This configuration retains the reverse physical path and removes the forward path.

(7) Raw adjacency. This configuration retains the complete physical backbone and uses only the original physical adjacency for propagation.

(8) D²STGNN Physical. This configuration replaces CANDOR's entire physical backbone with the physical modeling structure of D²STGNN.

4. Overall evidence from the ablations. The ablations show that the four-role decomposition contributes across most evaluated settings, while functional relations also provide useful predictive information in most settings. In summary, the functional and physical branches form complementary views that allow the same architecture to adapt to different domains and retain competitive forecasting performance.

Question3:

1. Dynamical assumptions. Appendix A.1 of the revised manuscript provides the detailed derivation of the dynamical assumptions motivating the four roles and the two background references.

2. Capacity-matched alternative decompositions. To evaluate the proposed decomposition, we compare it with four alternative decompositions under matched model capacity and the same forecasting framework:

  • Moving Average.
  • STL.
  • TimeMixer-style.
  • Trend-Season (DLinear-style).

Under matched model capacity, CANDOR achieves the lowest mean primary error on all five datasets. The detailed results and experimental specifications are provided in Appendix A.11 of the revised manuscript.

We hope the clarifications and evidence provided in these three comments address your Ethics Concerns. We thank you for the suggestions, which helped us clarify implementation attribution, strengthen the controlled comparisons, and explain the decomposition rationale more clearly.

Reviewer_Ue5LOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CANDOR is a forecasting framework for networked spatio-temporal systems. Its central premise is to first disentangle local dynamics and then learn the relations among nodes. Specifically, each node's historical trajectory is decomposed into four components—background, accumulation, release, and shock—and two relational branches are constructed conditioned on these components. The physical branch models edge-specific and sample-specific delayed propagation over a directed topology while preserving both forward and reverse directions, whereas the functional branch learns topology-agnostic non-local dependencies from the background dynamics. A context-adaptive fusion module then combines the three prediction streams: functional, forward, and reverse. The training objective comprises a main forecasting loss, branch auxiliary losses, a reconstruction loss, role consistency, historical consistency, a separation regularizer, background smoothness, and shock sparsity. Experiments cover two traffic benchmarks, METR-LA and PEMS-BAY, together with three newly compiled five-year river dissolved-oxygen datasets; the latter comprise 43 stations, approximately 1.88 million station-hours, and 87 directed river-network edges. The evaluation further compares 15 baselines, with the top eight reported in the main table, over three random seeds, and additionally includes ablation, parameter-sensitivity, complexity, and mechanism analyses. In the mechanism analysis, rainfall serves as an external probe for shock, and the fusion weights and non-local relations are also examined.

Critical strengths

The paper compares against 15 baselines with 3 random seeds on 2 traffic benchmarks and 3 newly constructed river dissolved-oxygen datasets. The model has only about 0.9M parameters, yet attains the best mean on every reported main metric. The water-quality data are also valuable in their own right: 43 stations, approximately 1.88 million station-hours, and directed river-network information. This convinces me that the results are not a coincidence on a single benchmark, and it further indicates that the method is viable for practical deployment. This is the primary reason I am willing to recommend a weak accept. In addition, the physical branch handles delay and direction, whereas the functional branch handles non-local relations under background conditions; the two are conceptually complementary. Ablations further show that different systems draw on different branches: traffic relies more on physical propagation, while functional relations matter more in rivers. Appendix A.10 additionally uses rainfall—excluded from training—as an external probe and performs RainAdd/RainOnly interventions. This makes the paper more than "yet another graph predictor"; it advances mechanistic hypotheses that can be tested further. The rainfall probe is particularly commendable, although the effect size still needs to be strengthened.

Critical weaknesses

The "w/o Decomp." ablation removes the decomposition module together with its parameters and inductive bias, but it reports no parameter- or compute-matched control and does not compare against standard trend–seasonality decompositions (moving average, STL, or TimeMixer-style decomposition) within the same architecture. Consequently, the 1.6% increase in METR-LA MAE and the 5.8% increase in BASIN180102 MSE cannot be attributed to the four-component decomposition alone. More problematically, in the authors' own ablations the degradation from removing decomposition is smaller than that from removing the physical branch (METR-LA MAE +9.5%) or the functional branch (BASIN180102 MSE +12.8%). This is difficult to reconcile with the title and abstract, which position decomposition as central to the contribution. Nor do the four components admit ground truth against which identifiability could be verified. The reconstruction loss (Eq. 5) and role consistency (Eq. 31) guarantee only additive consistency and soft alignment. The alignments in Figure 5(c) are precisely what L_role optimizes, so they cannot serve as independent evidence. The paper's central contribution is the principle of "decomposition before dependency learning," not any particular architecture; if the gains derive mainly from branch capacity or architecture, the claim at the level of principle is not sustained.

Questions for authors

Q1. Could the authors report the parameter count, FLOPs, and inference latency for all ablation variants? Two additional controls are also needed: first, a parameter-matched "w/o Decomp." variant, for example by widening or adding layers to the patch encoder until its parameter count is comparable to that of the full model; second, replacing the four-component decomposition with a standard trend–seasonality decomposition within the same two-branch architecture. Only then can the contribution of decomposition itself be separated from that of capacity and architecture. Q2. The water-quality task adopts MSE as its primary metric, yet the checkpoint criterion in Table 4 is validation MAE. Could the authors report results for a control in which model selection uses validation MSE? Is this choice applied symmetrically to CANDOR and the baselines? Q3. The Spearman ρ values in Table 13 are only 0.035/0.045/0.131. Although the p-values are significant, the effect sizes are small, and RainAdd yields no distinguishable improvement. Could the authors provide confidence intervals, per-horizon and per-basin results, and an additional negative control such as randomly shuffled or temporally shifted rainfall? How do the authors distinguish between "shock captures rainfall-related perturbations" and the possibility that rainfall is simply weakly related to this variable at this temporal scale? Q4. In Table 10, HyperD attains an MSE of 2.318 on BASIN030501, far above the 0.29–0.98 range of the other methods. Is this an implementation or tuning issue, or is the method itself unsuitable? If the former, should it be re-tuned, or removed from the "strongest baselines"?

5 comments
AuthorsOpenReview ↗

Weakness: Decomposition, parameter-matched controls, and semantic evidence

Weakness:

1. Parameter-matched w/o Decomp. We added a parameter-matched w/o Decomp. control on all five datasets, adjusting forecast-stage capacity so that its trainable parameter count differs from CANDOR by less than 0.08% using three seeds. Traffic is evaluated by Overall MAE and water quality by MSE at the 192-hour horizon. The mean errors are:

MethodMETR-LAPEMS-BAYBASIN030501BASIN031300BASIN180102
CANDOR2.8541.5080.6110.4871.542
Parameter-matched w/o Decomp.2.9001.5730.6550.4961.597

CANDOR retains lower mean error across all five datasets after capacity matching. Appendix A.11 provides the matching procedure, parameter counts, and complete mean and standard deviation results.

2. Alternative decompositions. Appendix A.11 also compares parameter-matched Moving Average, STL, TimeMixer-style, and Trend-Season (DLinear-style) configurations. CANDOR achieves the lowest mean error on all five datasets. Its reductions relative to the strongest alternative on each dataset are 1.37%, 2.97%, 2.80%, 0.90%, and 0.63%, in the dataset order above. Method definitions and full results are included in the revised Appendix A.11.

3. Explanation for why can branch removal have a larger effect. Decomposition and relation learning are closely connected in CANDOR. The physical branch separates local temporal states and propagation residuals, applies Shock modulation, and aligns these states and graph contributions with Background, Release, and subsequent Accumulation. The functional branch uses Background to learn cross-node relations. Decomposition therefore organizes both the representations supplied to the branches and their subsequent processing.

The branch ablations reflect the different information available in each system. Removing the physical branch raises METR-LA MAE by 9.5%, whereas removing the functional branch changes it by only 0.4%. On BASIN030501, removing the functional branch raises MSE by 6.5%, while removing the physical branch changes it by 0.7%. The smaller changes fall within the reported threshold based on seed variation. These differences motivate complementary physical and functional views whose contributions adapt across systems. Removing decomposition increases the primary error on all five datasets, and all five increases exceed the threshold based on twice the full-model seed variation in Appendix A.10. Its contribution is consistently observed across domains, while the relative importance of each branch varies with the system.

4. Semantic evidence beyond the optimized alignments. Figure 5(c) examines the correspondence learned between component representations and physical-path states. We have added a separate semantic analysis in Appendix A.12 using timestamp-defined traffic regimes. With the trained model fixed, we visualize Z(input embedding), A(Accumulation), and R(Release) for the same 1,200 test samples, balanced across nighttime, morning peak, and evening peak periods. The t-SNE displays on METR-LA and PEMS-BAY show clearer nighttime/peak separation in A and R than in Z, with substantial overlap between the two peak periods. This associates the component representations with recurring operating conditions through an evaluation criterion supplied by observation timestamps. The added figures and analysis protocol are included in Appendix A.12.

5. Relationship between the principle and architecture. The four components are reference-anchored latent roles motivated by local evolution and network transport, as developed in Appendix A.1. The architecture specifies how these roles guide dependency learning that Background informs functional relations, release-aligned residuals and Shock inform physical propagation, and graph contributions support subsequent Accumulation. The added capacity controls and alternative-decomposition comparisons test whether comparable capacity or conventional temporal components can reproduce the benefits of this role-conditioned organization. This connection between decomposition and downstream relation learning is the contribution expressed by “decomposition before dependency learning.”

AuthorsOpenReview ↗

Question1: Ablation complexity and parameter-matched decomposition controls

Question1:

1. Parameters, FLOPs, and inference latency. We profiled CANDOR, all ten ablation variants, and the five added parameter-matched controls on all five datasets. Traffic uses the 12-step task. Water quality uses H=192. All measurements use FP32 on the same RTX 5090, with batch size 64, 20 warm-up passes, and 100 timed passes.

Each cell below is parameters (M) / GFLOPs per input window / latency (ms per batch). FLOPs count matrix/convolution operations, including recurrent, attention, and graph matrix products, with a multiply-add counted as two operations. Scalar activations, normalization, softmax, top-k, and indexing are excluded. Latency measures the unchanged target-free forward pass, including graph construction and returned auxiliary diagnostics. Data loading and inverse transformation are excluded.

VariantMETR-LAPEMS-BAYBASIN030501BASIN031300BASIN180102
CANDOR0.899/2.758/44.520.904/2.953/43.420.877/0.460/46.470.878/0.769/47.890.877/0.422/45.98
w/o Decomp.0.289/0.911/7.850.307/1.228/9.130.267/0.122/13.790.268/0.205/14.000.267/0.112/13.62
w/o Aux. loss0.892/2.747/43.930.898/2.939/42.730.871/0.460/46.150.872/0.768/47.180.870/0.421/45.81
w/o Physical0.731/1.872/11.360.722/1.728/11.030.732/0.344/11.100.732/0.573/11.560.732/0.315/10.89
w/o Functional0.395/2.254/37.590.413/2.496/34.340.373/0.385/44.400.374/0.644/45.650.372/0.353/44.71
Forward only0.860/2.394/28.890.866/2.534/28.790.838/0.408/30.100.839/0.681/30.610.838/0.374/29.68
Reverse only0.860/2.394/28.830.866/2.534/28.780.838/0.408/29.530.839/0.681/30.790.838/0.374/29.99
Raw adjacency0.847/2.732/44.110.838/2.896/42.760.848/0.460/45.790.848/0.768/46.500.848/0.421/46.11
D²STGNN Physical0.842/2.652/18.510.840/3.008/19.420.833/0.402/17.970.834/0.673/18.870.833/0.368/18.27
w/o Relation0.832/2.513/41.310.850/2.738/40.160.810/0.436/45.030.811/0.729/46.330.810/0.400/46.03
Average fusion0.872/2.710/44.110.878/2.909/43.080.850/0.450/46.620.852/0.753/47.080.850/0.413/45.56
Matched w/o Decomp.0.899/2.669/12.320.904/2.770/13.090.877/0.486/14.500.878/0.812/15.240.877/0.445/14.60
Matched Moving Average0.899/2.687/13.840.905/2.824/14.950.877/0.482/14.860.878/0.805/15.760.876/0.442/14.87
Matched STL0.899/2.687/13.670.905/2.825/14.680.877/0.482/14.900.878/0.805/15.840.876/0.442/14.91
Matched TimeMixer-style0.899/2.688/14.290.905/2.825/14.790.877/0.482/15.120.878/0.805/15.930.876/0.442/14.95
Matched Trend-Season0.899/2.687/13.880.905/2.825/15.040.877/0.482/14.830.878/0.805/15.700.876/0.442/14.78

2. Parameter-matched controls. We added a parameter-matched w/o Decomp. and four parameter-matched alternatives including Moving Average, STL, TimeMixer-style, and Trend-Season (DLinear-style). Forecast-stage capacity is increased before fusion while retaining each control's processing. Every added configuration is evaluated with three seeds. Parameter counts differ from CANDOR by less than 0.08%. Their FLOPs range from 6.19% below to 5.55% above CANDOR.

The table reports Overall MAE for traffic and pooled MSE at H=192 for water quality, as means ± sample standard deviations:

MethodMETR-LAPEMS-BAYBASIN030501BASIN031300BASIN180102
CANDOR2.854 ± 0.0091.508 ± 0.0080.611 ± 0.0090.487 ± 0.0061.542 ± 0.011
Parameter-matched w/o Decomp.2.900 ± 0.0061.573 ± 0.0040.655 ± 0.0080.496 ± 0.0011.597 ± 0.015
Moving Average2.895 ± 0.0031.555 ± 0.0030.629 ± 0.0040.500 ± 0.0101.642 ± 0.047
STL2.898 ± 0.0031.554 ± 0.0060.630 ± 0.0110.492 ± 0.0071.624 ± 0.058
TimeMixer-style2.893 ± 0.0101.554 ± 0.0030.632 ± 0.0150.500 ± 0.0041.607 ± 0.018
Trend-Season (DLinear-style)2.901 ± 0.0061.554 ± 0.0030.631 ± 0.0200.500 ± 0.0061.552 ± 0.013

CANDOR retains the lowest mean primary error on all five datasets. The revised Appendix A.11 provides the decomposition definitions, matching procedure, parameter counts, and forecasting results.

AuthorsOpenReview ↗

Question2: Sensitivity to the checkpoint-selection metric

Question2:

1. Checkpoint-selection rationale. MAE and MSE are both established forecasting metrics. CANDOR uses validation MAE to keep checkpoint selection consistent across traffic and water-quality tasks, and MAE is also a reported water-quality evaluation metric. Several baseline implementations likewise use validation MAE.

2. Retrospective selection comparison. For runs whose original checkpoints were selected using validation MAE, we compared the reported results with those obtained by selecting checkpoints from the same training logs using validation MSE or RMSE. Recorded MSE and RMSE criteria are included in this sensitivity analysis. Training remains unchanged, and ties in the logged minimum are resolved by selecting the first occurrence.

Each cell below reports three relative changes in test MSE, ordered as BASIN030501 / BASIN031300 / BASIN180102. Changes are percentages, computed as 100 × (alternative-selection mean MSE / original-selection mean MSE − 1). The original-selection results are those used in the manuscript. HyperD uses its original configuration.

Method48 h96 h144 h192 h240 h
CANDOR-0.16/+0.18/-1.39-0.73/+0.71/-0.25-0.40/+0.17/+0.19+0.63/+0.39/0.000.00/+0.46/-0.21
D²STGNN-0.35/0.00/-0.120.00/-0.43/-1.64-0.23/+0.55/+0.250.00/0.00/+1.060.00/-4.35/+2.81
PDFormer-0.77/+0.24/+0.540.00/0.00/0.00-0.50/-0.51/0.000.00/+0.03/0.000.00/0.00/0.00
HyperD0.00/0.00/-0.150.00/0.00/+0.39+0.60/+0.27/0.000.00/0.00/0.000.00/+0.26/+1.99
RSTIB-MLP0.00/+0.09/+0.08-0.01/-0.01/+0.01+0.29/+0.01/0.00-0.05/-0.07/+0.14-0.02/-0.16/0.00
STPGNN-0.46/-0.01/+0.08-2.10/0.00/+0.34+1.23/-0.04/+3.60-0.08/-1.89/+0.01-0.04/+0.06/+6.13
TESTAM0.00/-2.59/0.00+0.23/0.00/-0.64-1.54/+1.91/-5.45+2.76/0.00/-1.030.00/+2.68/0.00
LatentTSF0.00/0.00/-0.16-0.24/+0.03/-0.26-0.42/0.00/-0.38-0.44/0.00/-0.15-0.73/0.00/-0.32

3. Sensitivity of the results. Across the 120 method-basin-horizon comparisons, 102 (85.0%) change by at most 1% in absolute value, and 111 (92.5%) by at most 2%. CANDOR's changes range from -1.39% to +0.71%. It retains the lowest five-horizon mean MSE among the listed methods on all three basins under both selections.

The records show limited sensitivity to the selection metric in most settings and preserve the main comparison.

Reviewer_ERq9OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces CANDOR, a method which decomposes local dynamics in spatiotemporal systems before attempting to learn dependencies. The method embeds node histories into patches and constructs four learned components (background, accumulation, release, and shock). The components are then recombined into a representation used by two relation mechanisms: a topology-constrained physical branch with forward/reverse, temporal-position-dependent aggregation, and a functional branch whose relation scores depend on background descriptors. Experimental results cover two traffic systems (METR-LA, PEMS-BAY), and three river water quality datasets, with forecasting horizons up to 240 hours. Consistent improvements over many competitive baselines under shared evaluation settings are observed, and model ablations help reveal contributions of each mechanism.

Critical strengths
  1. The paper's stated contributions are clear and attempt to improve performance in the challenging domain of spatiotemporal modeling. The introduced approach has a coherent predictive rationale and nicely includes an external interpretability check. The separation of persistent background from transient variation before selecting relations appears reasonable, especially when physical connectivity alone is incomplete. There are additionally helpful analyses like presented in Appendix A.10, along with transparent acknowledgement of method boundaries, e.g., that the components do not obey exact physical conservation.
  2. The empirical evidence is quite comprehensive and well-presented. Two systems with different temporal scales and graph structures are used for evaluations, covering variations in the underlying domain dynamics. Identical splits and evaluation pipelines are used when comparing methods, and three-seed results are reported for all methods in Appendix A.8. There is reasonably strong performance improvement across metrics on each dataset, and ablations extend across all five datasets, making it easier to judge any setting-specific model behaviors across components. The method is also quite compelling for broader practice use give its fairly small size and profiling on common/consumer hardware.
Critical weaknesses
  1. There are a few key related works/baselines whose absence appears to undermine the core novelty claims. Most notably, D2STGNN is similar in that it separates inherent and diffusion information with an estimation gate/residual decomposition, and the separated signals are processed with specialized modules and dynamic graph learning. Its focus on the traffic systems make it an especially strong baseline candidate, yet it's absent from both the related work discussion and the benchmark evaluations. The CANDOR method's four distinct roles and background-conditioned functional relations are potentially meaningful differences, but the broad claim that prior methods generally learn a single dependency structure from entangled observations appears too general. PDFormer also seems like a possibly strong baseline candidate which explicitly addresses propagation delay, and while discussed it's also not evaluated.
Questions for authors
  1. What distinguishes CANDOR from D2STGNN at the level of information separation and branch specialization, and how do they compare under the same forecasting protocol? Within CANDOR's backbone, does the proposed four-component construction outperform a capacity or loss-matched generic decomposition or trend/residual split? This would help determine whether the central method design adds more than useful architectural complexity and strengthen its stance relative to a similar piece of literature.
  2. As far semantic validation, is there evidence independent of the role losses that accumulation/release track with their intended processes? The same would go for whether the proposed decomposition improves recovery of known relation/delay structures. The rainfall probe is revealing for shock, but analogous evidence for the main propagation interpretation would be particularly insightful here.
4 comments
AuthorsOpenReview ↗

Weakness and Question1: Baseline coverage and comparison with new benchmarks and decomposition methods

Weakness:

1. Expanded baseline comparisons. Our original set of 15 baselines prioritized representative methods published at major venues within the preceding three years, to reflect recent advances in forecasting. D²STGNN (2022) and PDFormer (2023) fell outside this selection window. We have now added both methods, expanding the comparison to 17 baselines. They use the same chronological splits, input histories, forecast horizons, targets, and original-scale evaluation. The complete results are provided in Section 4.1 and Appendix A.9 of the revised manuscript.

2. Related work and physical-module replacement. We have added an explicit discussion of D²STGNN in Related Work and revised the overly broad characterization of previous dependency-learning methods. We also added an ablation replacing CANDOR’s entire physical module with a D²STGNN decoupling layer, while retaining the upstream decomposition, patch grid, and functional branch. The replacement increases METR-LA MAE by 1.0% and water-quality MSE by 1.1%, 0.3%, and 6.6% across the three basins at the 192-hour horizon, with comparable performance on PEMS-BAY. These results are reported in Appendix A.10.

3. Cross-system modeling objective. These additions strengthen the empirical basis for CANDOR’s contribution. Our objective is to develop a shared framework that performs effectively across domains with different dynamics and interaction structures. The expanded comparisons and replacement experiment support this objective across urban traffic and river water quality. CANDOR’s complementary physical and functional mechanisms allow the same architecture to exploit different sources of predictive information across systems, supporting its broader applicability to spatiotemporal forecasting.

Question1:

1. Information separation and branch specialization. D²STGNN uses an estimation gate and residual decomposition to separate Diffusion and Inherent signals. Its diffusion module models cross-node propagation, while its inherent module captures node-local temporal evolution.

CANDOR organizes local dynamics into Background, Accumulation, Release, and Shock before learning dependencies. These roles are motivated by local state continuation, incoming contributions, outgoing release, and disturbances. The current and subsequent historical Background references provide distinct anchors for Accumulation and Release. Appendix A.1 of the revised manuscript explains the dynamical assumptions motivating this construction.

The branch specialization also differs. D²STGNN separates cross-node diffusion from local temporal evolution. CANDOR separates physical propagation relations from functional associations. Both major branches in CANDOR use cross-node information. The Physical Branch models forward propagation and reverse-support interactions over directed physical connections. The Functional Branch uses Background to discover predictive dependencies beyond physical adjacency. Consequently, CANDOR’s branches serve different purposes from D²STGNN’s branches.

2. Comparisons under the same forecasting protocol. We added D²STGNN as a full-model baseline using the same chronological splits, input histories, forecast targets, horizons, and original-scale evaluation, without additional forecasting covariates. CANDOR achieves lower mean Overall MAE on both traffic datasets: 2.854 versus 2.870 on METR-LA and 1.508 versus 1.514 on PEMS-BAY. Across the three water-quality datasets, its five-horizon mean MSE is 15.24-17.00% lower than D²STGNN. Complete results are provided in Section 4.1 and Appendix A.9.

We also replaced CANDOR’s entire physical module with a D²STGNN decoupling layer while retaining the upstream decomposition, patch grid, and Functional Branch. The replacement increases METR-LA MAE by 1.0% and water-quality MSE by 1.1%, 0.3%, and 6.6% across the three basins at the 192-hour horizon, with comparable performance on PEMS-BAY. Appendix A.10 reports this comparison.

3. Capacity-matched decomposition controls. We added a parameter-matched w/o Decomp. control and Moving Average, STL, TimeMixer-style, and Trend-Season (DLinear-style) alternatives across all five datasets. Parameter counts differ from CANDOR by less than 0.08%, with shared prediction, branch-supervision, and historical-consistency losses retained. Each added control uses three seeds.

CANDOR achieves the lowest mean primary error on all five datasets. Relative to the strongest alternative decomposition, the reductions are 1.37% on METR-LA, 2.97% on PEMS-BAY, 2.80% on BASIN030501, 0.90% on BASIN031300, and 0.63% on BASIN180102. The parameter-matched w/o Decomp. also remains worse across all five datasets. Detailed configurations, parameter counts, and complete results are provided in Appendix A.11.

AuthorsOpenReview ↗

Question2: Semantic validation of components and explanation of propagation relations

Question2:

1. Semantic evidence independent of the role-alignment objective. Considering meaningful Accumulation and Release representations should reflect differences in the intensity of vehicle buildup and discharge, we compare low-activity nighttime conditions with morning and evening peak periods to examine whether these differences are reflected in the learned feature space. We added an analysis of traffic-regime structure in the learned Accumulation and Release representations on METR-LA and PEMS-BAY. With the trained models fixed, we extract Z(input embedding), A(Accumulation), and R(Release) for the same test samples and visualize them using t-SNE. The samples cover three timestamp-defined operating conditions: weekday nighttime, morning peak, and evening peak. Each display contains 400 samples per condition, with the same 1,200 samples used across representations.

A and R exhibit clearer separation between nighttime and peak-period samples than Z. Moreover, morning and evening peak samples substantially overlap, consistent with characteristics shared by the two peak periods. These timestamp-defined operating conditions provide an evaluation criterion separate from the component-to-state alignments optimized by the role loss. The results connect the learned representations to real traffic conditions and complement Figure 5(c) and indicate that the learned Accumulation and Release representations capture meaningful semantic differences between real-world traffic operating conditions. The corresponding figures and protocol are provided in Appendix A.12.

2. Correspondence between temporal-position weights and road distance. We also examined the relationship between road distance and temporal-position weighting among the neighbors of the same receiving node. The analysis uses the complete METR-LA test set of 6,850 windows and compares Kt=2K_t=2 and Kt=3K_t=3.

For each edge, we summarize its preference over all positions in the localized historical window using a weighted average of normalized position age with the earliest position age 1 and the most recent position age 0. Within each receiving node, we compare this quantity with road distance and compare the longer-distance and shorter-distance neighbor groups. The Forward-path results on the same 168 eligible receiving nodes are:

QuantityKt=2K_t=2Kt=3K_t=3
Nodes whose longer edges favor earlier positions more than shorter edges48.8% (82/168)63.1% (106/168)
Median within-node Spearman correlation between distance and position age−0.02380.1642

For Kt=3K_t=3, the positive relationship is consistent with the expectation that longer propagation paths can require more time under comparable conditions. The relationship is weaker for Kt=2K_t=2, although that configuration has better forecasting performance.

Interpreting learned spatiotemporal weights requires considering several coupled factors. Road distance captures one aspect of propagation, while traffic conditions, time of day, and local network structure can also affect effective interactions. A direct semantic correspondence between an individual factor and the final learned weights is difficult to establish. Ground-truth interaction weights and propagation-delay labels are unavailable in these datasets.

The traffic-regime analysis shows that the decomposition extracts role representations with meaningful structure associated with observed operating conditions. Such semantic organization provides a useful prior for subsequent propagation and functional-relation learning, whose predictive benefit is supported by the matched-capacity comparisons and cross-dataset ablations. Further separating the effects of path length, traffic conditions, and temporal context would help establish a more explicit causal interpretation of the learned temporal-position weights.

We thank you for the questions and suggestions, which helped us expand the baseline comparisons, clarify CANDOR's relationship to D²STGNN, and strengthen the semantic analysis of the learned components and propagation weights.

I'd like to thank the authors for their detailed and insightful reply, and the willingness to engage with feedback across the board is encouraging. I remain positively inclined toward acceptance and will therefore keep my score.

AuthorsOpenReview ↗

Thank you for your careful review

Thank you for your careful review and constructive feedback. We appreciate your positive assessment and continued support for our work.

Comments on the paper

2 comments

Dear Reviewers,

As the discussion period draws to a close, please read and acknowledge the authors’ rebuttal within the next 24 hours. I also encourage you to engage in the discussion and share any remaining questions or concerns.

Thank you for your time and thoughtful reviews.

Best,

Your AC

AuthorsOpenReview ↗

Acknowledgments and Summary of Revisions

Dear Area Chair and Reviewers,

Thank you for your careful reading and constructive feedback throughout the review process. We also thank the Area Chair for coordinating the discussion and encouraging further engagement. Your comments have helped us sharpen CANDOR’s contributions and strengthen the supporting evidence.

  • Reviewer ERq9: Your questions about baseline coverage and independent semantic validation motivated broader comparisons and additional analyses of traffic-regime representations and temporal-position weights. These additions help clarify both the meaningful structure captured by the components and the limits of their physical interpretation.
  • Reviewer Ue5L: Your suggestions led to parameter-matched decomposition controls, detailed computational profiling, checkpoint-selection sensitivity analysis, and stronger rainfall controls. They also motivated further baseline adaptation experiments, making the evaluation more comprehensive and transparent.
  • Reviewer EtYX: Your close examination of the method and implementation helped us articulate the architectural innovations more precisely and explain how decomposition supports subsequent relation learning. Your questions also motivated matched comparisons and a clearer account of the dynamical assumptions underlying the four roles and dual background references. We appreciate your continued engagement and reassessment.

These revisions clarify how CANDOR connects local dynamical roles with complementary relational mechanisms and strengthen the evidence for its effectiveness across two distinct spatiotemporal systems.

We hope the revised manuscript and responses address the key concerns. Thank you again for your time and thoughtful consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.