acceptodds
Under review as a conference paper at ICLR 2027

Learning Networked Dynamical Systems via Neural ODE-based Graph Structure Inference

Abstract

Inferring the latent interaction graph from multivariate time series is fundamental to understanding complex dynamical systems in neuroscience, biology, and physics. Despite recent progress in neural relational inference, most methods rely on discrete-time modeling, which degrades under sparse and irregular observations and tends to entangle multi-hop dependencies into spurious direct edges. We show that a continuous-time formulation naturally decouples inference from any fixed observation schedule, allowing gradient signals to be integrated continuously between observations for both dynamics and structure learning. We propose NOGI (Neural ODE-based Graph Inference), a framework that couples a deterministic graph learner with an MPNN-based Neural ODE. The two modules are trained end-to-end via the adjoint method with memory cost. Across four simulated dynamical systems at two scales, NOGI achieves the best structure recovery in 30 out of 32 evaluated cases with highly competitive forecasting performance, remaining robust under observation degradation. On the realistic Netsim fMRI benchmark, it attains an average AUPRC of 0.90 (AUROC 0.96). These results suggest that explicitly decoupling relational interactions from self-dynamics within a continuous-time vector field provides a powerful mechanism for graph structure learning under realistic observation conditions.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph structure learning, neural relational inference, complex network reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_V9urOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces NOGI, a framework for learning network structure and continuous-time dynamics together from observed trajectories of the studied system. NOGI combines an explicitly learned sparse adjacency matrix with a graph-conditioned Neural ODE to model system dynamics. Experiments on several synthetic dynamical systems and the Netsim benchmark demonstate strong structure-recovery performance, including under sparse and irregular observations.

Critical strengths
  1. The joint learning of network structure and continuous-time dynamics is well motivated and potentially useful for systems with irregular or incomplete observations. But this idea is not completely new, e.g., GNeuralFlow (NeurIPS 2024) also jointly learns graph structure and continuous-time dynamics from irregularly sampled time series.

  2. The evaluation in the paper covers multiple dynamical systems, graph topologies, network sizes, and missing-data regimes. The reported AUROC and AUPRC results are generally strong, and the paper investigates challenging observation settings as well.

Critical weaknesses
  1. Although the experiments conducted in the paper are broad, the baseline comparisons do not establish comprehensive coverage of recent structural-inference methods. Comparisons with approaches such as GNeuralFlow, iSIDG, SIDEC would help clarify NOGI’s relative contribution.

  2. The performance of NOGI is not always superior across all settings, so the scope of the paper’s empirical superiority claims should be carefully qualified.

  3. It is difficult to isolate the benefits of the continuous-time formulation and joint graph-dynamics learning without sufficiently targeted comparisons or ablations.

Questions for authors

Related to the identified weaknesses:

  1. Can the authors compare NOGI with additional structural-inference methods, ideally on at least one shared benchmark with consistent data splits and AUROC/AUPRC metrics?

  2. Which components of NOGI contribute most to its performance, and how do the authors disentangle their effects?

2 comments
AuthorsOpenReview ↗

Response to Reviewer V9ur

W1/Q1:

We compare NOGI with iSIDG and SIDEC on four datasets from the public StructInfer benchmark. Baseline numbers are taken from Table 1 of the SIDEC paper, and NOGI follows the same protocol: the official data splits, AUROC over the full adjacency matrix, and an average over 10 runs. NOGI outperforms iSIDG on all four datasets, outperforms SIDEC on VN_SP_50 and VN_SP_100, and is slightly behind SIDEC on VN_NS_50 and VN_NS_100.

AUROCVN_SP_50VN_SP_100VN_NS_50VN_NS_100
iSIDG0.96±0.020.85±0.020.74±0.020.69±0.02
SIDEC0.97±0.030.96±0.020.99±0.020.99±0.01
NOGI0.99±0.011.00±0.000.98±0.040.98±0.06

W2:

We would like to clarify that our claims are explicitly scoped to the evaluated settings: we report best or jointly best structure recovery in 30 of 32 comparisons, explicitly identify the exceptions in §3.2, and discuss settings where NOGI does not outperform competing methods in §§3.3–3.4.

W3/Q2:

We ablate one component at a time on Spring and Kuramoto (ER, N=100) under observation interval K, with all other settings fixed within each column. Each cell reports AUPRC over three seeds:

VariantSpring K=1Spring K=50Kuramoto K=1Kuramoto K=10
NOGI1.00±0.000.95±0.021.00±0.000.89±0.20
w/o continuous time1.00±0.000.85±0.011.00±0.000.83±0.29
w/o sparsity0.90±0.160.95±0.021.00±0.000.87±0.21
w/o unbounded W0.70±0.260.51±0.121.00±0.000.41±0.06
w/o deterministic W0.34±0.260.40±0.040.84±0.110.42±0.04
w/o pairwise message1.00±0.000.54±0.400.04±0.010.04±0.01

The benefit of continuous-time modeling appears at large intervals: over a long interval, influence spreads across multiple hops, which a discrete-time model can only explain with spurious direct edges, whereas NOGI's graph describes only instantaneous interactions and lets the integration produce the multi-hop effects. For the other components, deterministic graph learning contributes most, pairwise message passing is essential on Kuramoto, and sparsity regularization gives a further gain. We have added this ablation study to the revised paper. As for joint graph-dynamics learning, it is the task setting shared by NRI, AIDD, NGM, and SCOTCH rather than a design choice of NOGI, and Table 1 already contrasts this setting with methods that infer the graph without a dynamics model (MI, PC).

Thank you for the response which (partially) addresses the weaknesses from my review. I also look at the other reviews, author responses and discussions. I can raise my score a bit, but still feel that the work does not meet the standards of ICLR conferences.

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

The paper proposes NOGI for jointly learning interaction graphs and continuous-time dynamics from multivariate time series. It combines a directly optimized, nonnegative adjacency matrix with a shared message-passing neural ODE, trained using prediction loss, sparsity regularization, and adjoint sensitivities. Evaluation covers four simulated dynamical systems, observational degradation, transfer across graph topologies and sizes, and 27 Netsim configurations. The reported results show strong graph recovery, particularly at 100 nodes, alongside competitive forecasting performance.

Critical strengths

Strong reported structure-recovery performance across varied settings.

Table 1 reports leading results in 30 of 32 AUROC/AUPRC comparisons, with particularly strong performance at 100 nodes. Extending evaluation to 27 Netsim configurations also provides useful coverage beyond a single commonly studied subset. These results make the proposed combination empirically promising.

A simple architecture with useful evidence of transferability.

Explicit edge weights and shared message-passing functions provide a straightforward way to combine graph inference with continuous-time prediction. Table 3 shows that a frozen dynamics module can support graph recovery on different topologies and sizes after optimizing the target graph. This is a useful practical result, even though it does not establish recovery of the true dynamical rules.

Critical weaknesses
  1. The paper does not sufficiently establish its explanation for the improvements.

The distinction between instantaneous interactions and finite-interval predictive dependencies is important, but the introduction and Fig. 1 blur this distinction with numerical approximation accuracy. A learned discrete-time transition model is not necessarily a first-order Euler approximation; even an exact transition map can contain multi-hop dependencies absent from the instantaneous graph. Conversely, higher-order integration alone does not guarantee correct graph recovery.

The attribution in §3.2 to an “explicit graph decoupled from the dynamics” also needs a sharper distinction from prior work. SCOTCH already infers a separate graph variable and explicitly conditions its neural dynamics on it (Eqs. (8)–(11)). Relative to AIDD, NOGI changes both the temporal formulation and graph-learning procedure; relative to NGM, which already uses a Neural ODE, it changes the graph representation, regularization, and vector-field architecture. Table 2’s unstructured-NODE and ground-truth-graph controls do not isolate these differences. Table 3 supports reuse of the learned dynamics, but does not establish that structural decoupling explains the gains over competing methods.

Consequently, the experiments support the combined design more directly than the proposed attribution of its improvements. Comparative computational scalability also remains insufficiently evaluated: recovery accuracy at 10 and 100 nodes and constant adjoint memory with respect to solver steps do not establish runtime or memory advantages as graph size and sequence length increase.

  1. The evaluation protocol is insufficiently specified to assess the headline gains confidently

Appendix A.4 provides a useful ground-truth-free procedure for selecting the sparsity coefficient, but does not clearly establish whether this procedure was used for the main results or how checkpoints were selected. For example, the selected configuration in Table 6 achieves 0.89 AUPRC on Netsim Sim3, whereas Table 4 reports 0.93. These may reflect different configurations, but the relationship is unexplained.

The Netsim comparison also needs a precise account of subject selection or pooling, preprocessing, self-connection scoring, and baseline training. These details matter because the main empirical contribution is improved recovery of an unknown graph: readers need to know how models were selected without that graph and whether the methods were evaluated under matched conditions.

Questions for authors
  1. Model selection and evaluation protocol.

How were the sparsity coefficient, any decay schedule, and checkpoints selected for Tables 1 and 4? Was the stability criterion in Appendix A.4 used for these results, and what explains the Sim3 difference between Tables 4 and 6? Please also specify the train/validation/test split unit and whether the Netsim SCOTCH results are new runs using the same subjects, preprocessing, and self-connection scoring as NOGI.

  1. Evidence for the claimed source of improvement.

Is there a comparison that holds the message-passing architecture and graph-learning procedure fixed while varying the temporal formulation? Such a control would help determine whether the gains arise from continuous-time modeling or other design differences. What runtime or peak-memory evidence supports the claimed computational advantage as node count and sequence length increase?

  1. Scale ambiguity in the learned graph.

In Eqs. (5)–(6), let (B_m) denote the block of the update MLP’s first-layer weights acting on the aggregated message. The transformation (W\mapsto cW), (B_m\mapsto B_m/c), for (0<c<1), appears to preserve the vector field while reducing (\lambda\lVert W\rVert_1). Are there normalization constraints or additional regularizers that prevent this compensation? This does not itself invalidate ranking-based graph metrics, but it affects the interpretation of edge magnitudes and the role of the sparsity penalty.

  1. Interpretation of transferred dynamics.

Appendix A.2 uses heterogeneous node-specific frequencies in Kuramoto and inputs in FHN, while Eqs. (5)–(6) describe shared functions receiving only node states. Are these heterogeneous parameters supplied as additional inputs? If not, how do the authors distinguish recovery of transferable intrinsic dynamics from target-specific compensation through the learned graph? This clarification would affect the interpretation of Table 3 and the claim that the model captures the true underlying dynamics.

1 comment
AuthorsOpenReview ↗

Response to Reviewer pfse

We appreciate your detailed comments and the opportunity to clarify our method and experimental results.

W1/Q2:

We acknowledge that the original Fig. 1(b) blurred this distinction with numerical accuracy. NOGI's gain comes not from a higher-order solver but from what the graph is asked to explain: in discrete-time methods, it explains a finite-interval transition map, whose support contains multi-hop paths. In NOGI, it only parameterizes the instantaneous vector field, and multi-hop effects emerge from the integration, which keeps the graph free of spurious direct edges. This distinction was established by NGM (Bellot et al., 2022). We have revised Fig. 1(b) to clarify this distinction in the revised version.

To verify this, we ran the matched comparison requested in Q2: with the graph learner and message-passing architecture unchanged, we replace the ODE integration with a single discrete update that predicts the next observation directly.

AUPRCSpring K=1Spring K=50Kuramoto K=1Kuramoto K=10
NOGI1.00±0.000.95±0.021.00±0.000.89±0.20
w/o continuous time1.00±0.000.85±0.011.00±0.000.83±0.29

The advantage of continuous-time modeling is clear at a large interval. The full ablation has been added as §3.7 of the revised paper, and we agree that the attribution in §3.2 was too strong: it now points to this ablation instead.

On computational scalability, we measure NOGI on Spring (ER, dopri5) with adjoint and direct backpropagation. Each cell reports epoch time / peak memory / forward,backward NFE per epoch:

N, KAdjointDirect backprop
10, 11.1s / 43MiB / 168,1680.8s / 192MiB / 168,0
100, 15.4s / 1879MiB / 168,1683.5s / 15954MiB / 168,0
100, 101.2s / 1873MiB / 52,70OOM
100, 500.8s / 449MiB / 50,1040.3s / 13616MiB / 50,0
100, 1001.3s / 232MiB / 74,2240.3s / 10030MiB / 74,0

The adjoint method uses substantially less memory than direct backpropagation. Epoch time and memory are lower at larger intervals because subsampling leaves fewer training samples. We acknowledge that the "O(1) memory cost" stated in the abstract, §1, and §2.2 was imprecise: memory stays constant as the number of solver steps (NFE) grows, but it still grows with sequence length and network size. We have corrected the memory statements in the revised version.

W2/Q1:

λ in Table 1 was set per system from the grid. The stability criterion in Appendix A.4 was developed afterwards as a ground-truth-free check, and the λ it selects yields AUPRC close to the results in the main tables. For Table 4, a single λ = 1e-5 was used for all 27 simulations. λ was held constant during training. No decay schedule was used in the reported experiments, and we have removed the corresponding phrase from Appendix A.4. The reported graph is the one at the final training epoch. We have added these details and corrections to Appendix A.4.

The two values of Sim3 share the same configuration and differ only in the random seeds used for the three runs. On Sim3 the seed-to-seed AUPRC ranges from 0.85 to 0.94.

For all datasets, trajectories are segmented into single-step prediction samples, which are pooled and randomly split 60/20/20. The Netsim SCOTCH results are new runs with its official Netsim hyperparameters. In Table 4, NOGI and SCOTCH use exactly the same data, namely all 50 subjects of each simulation and the raw signals without normalization, and the same ground truth, except that SCOTCH's self-connection entries were scored as learned while NOGI's were set to zero. We have updated Table 4 in the revised version with results obtained by scoring both methods with the diagonal set to zero. The corrected SCOTCH row improves, and NOGI still leads on every simulation.

Q3:

There is no normalization constraint or additional regularizer that prevents this compensation. In practice, however, the L1 penalty acts by pruning uninformative edges and does not shrink the whole graph: on Kuramoto and Spring (ER, N=100), 67% and 45% of non-edges are driven exactly to zero while true edges keep an average weight of about 0.3. Accordingly, we interpret W only through its relative ranking (§3.1), not as absolute coupling strengths. We have clarified in §2.2.1 of the revised version that W represents relative edge strengths.

Q4:

These heterogeneous parameters are not supplied to the model, whose vector field receives only node states, so we expect node-specific effects to be absorbed by the learned W rather than by gθ, although we have no theoretical guarantee of this. Table 3 tests exactly this situation: the source and target systems share the same governing equations but have different Iᵢ, yet the frozen gθ still recovers the target graph. The transfer therefore holds not only across topologies and sizes but also under changed node-specific parameters.

Reviewer_wT1jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a NOGI, which jointly learns a deterministic nonnegative graph and an MPNN-based Neural ODE framework for multivariate signals. It combines a deterministic, nonnegative graph learner with an MPNN-based Neural ODE, trained through adjoint sensitivity using prediction loss and sparsity regularization. Experiments on four simulated dynamical systems and the Netsim fMRI simulation benchmark demonstrate strong graph recovery, competitive forecasting, and robustness under selected observation-degradation settings. The ground-truth-free regularization-selection analysis is also useful.

Critical strengths
  • NOGI integrates an explicit graph learner with an MPNN-based Neural ODE, providing a direct path from prediction errors to graph updates.

  • The method achieves the best or jointly best results in 30 of 32 reported metric comparisons and performs strongly across 27 Netsim simulations.

  • Comparisons include continuous-time methods such as NGM and SCOTCH, complemented by observation-degradation experiments, dynamics-transfer tests, and ground-truth-free regularization selection.

Critical weaknesses
  • ODE Integration starts directly from the observed snapshot rather than the learned latent states. This assumes the observation is a sufficient dynamical state. The implications of noisy initialization, partial observation, and hidden variables are not adequately evaluated.

  • Explicit graph parameterization does not establish structural disentanglement. Equation 6 uses a joint nonlinear update rather than the additive self/interaction decomposition. Compensating rescaling of edge weights and message outputs can also preserve predictions while changing the graph L1 penalty. Strong recovery results are encouraging, but do not establish unique graph dynamics identification.

  • External comparisons change multiple design choices simultaneously. Matched ablations are needed to separate the benefits of deterministic graph learning, message passing, sparsity regularization, and continuous-time integration.

  • Authors mention dopri5 or RK4 but do not fully specify choices, selection criteria, tolerances, or step sizes. Without solver and accuracy sensitivity analyses, numerical effects cannot be separated from architectural gains.

Questions for authors
  • Which gains remain when graph learning, vector-field capacity, loss, and compute are controlled independently?

  • What solver, step size, and tolerances are used for each system, and how sensitive are results to these choices?

  • Can forward/backward NFE, runtime, and peak memory be reported across network sizes and observation intervals?

  • How does performance change with noisy initial states or partially observed dynamics?

2 comments
AuthorsOpenReview ↗

Response to Reviewer wT1j

Thank you for your careful review and constructive feedback.

W1/Q4:

NOGI makes single-step predictions starting from the observed snapshot, so noisy initial states are exactly the setting of Fig. 3(a), where NOGI degrades most gracefully on Spring and MM, maintaining 0.84 AUPRC on Spring even at 5 dB.

We further evaluate NOGI under partially observed states and hidden nodes:

MetricSpring w/o vᵢFHN w/o wᵢSpring w/o 10% nodesSpring w/o 30% nodesFHN w/o 10% nodesFHN w/o 30% nodes
AUROC0.550.891.00±0.001.00±0.000.83±0.280.97±0.01
AUPRC0.050.521.00±0.001.00±0.000.66±0.530.69±0.04

For partially observed states, we omit the velocity vᵢ on Spring and the recovery variable wᵢ on FHN: recovery fails on Spring, but remains partial on FHN, where the coupling acts through the observed vᵢ. For hidden nodes, we randomly drop 10% or 30% of the nodes from the observations and recover only the edges among the remaining observed nodes, repeating with three random draws: recovery on Spring is robust in all draws, while one of the three draws for FHN w/o 10% nodes fails (AUPRC: 0.95, 0.97, 0.05).

W2:

By "decoupling" we mean that the graph and the dynamics are learned as separate components, so the dynamics module does not depend on a particular topology: in Table 3, a dynamics module trained on one topology transfers to other topologies and network sizes with only the graph relearned. The joint update in Eq. 6 is a more general form of the additive decomposition in Eq. 1. The disentanglement is supported empirically rather than guaranteed theoretically.

Rescaling the graph and the message outputs indeed preserves the predictions, but since the message network is shared by all edges, it can only multiply the whole graph by a common factor, leaving the edge ranking (AUROC/AUPRC) unchanged. Moreover, the learned W shows no sign of such rescaling: on Kuramoto and Spring (ER, N=100), 67% and 45% of non-edges are driven exactly to zero, the largest non-edge weight is below 0.1, and true edges keep an average weight of about 0.3. The L1 penalty thus prunes uninformative edges rather than shrinking the whole graph.

W3/Q1:

We ablate one component at a time on Spring and Kuramoto (ER, N=100) under observation interval K, with all other settings fixed within each column. Each cell reports AUPRC over three seeds:

VariantSpring K=1Spring K=50Kuramoto K=1Kuramoto K=10
NOGI1.00±0.000.95±0.021.00±0.000.89±0.20
w/o continuous time1.00±0.000.85±0.011.00±0.000.83±0.29
w/o sparsity0.90±0.160.95±0.021.00±0.000.87±0.21
w/o unbounded W0.70±0.260.51±0.121.00±0.000.41±0.06
w/o deterministic W0.34±0.260.40±0.040.84±0.110.42±0.04
w/o pairwise message1.00±0.000.54±0.400.04±0.010.04±0.01

The benefit of continuous-time modeling appears at large intervals: over a long interval, influence spreads across multiple hops, which a discrete-time model can only explain with spurious direct edges, whereas NOGI's graph describes only instantaneous interactions and lets the integration produce the multi-hop effects. For the other components, deterministic graph learning contributes most, pairwise message passing is essential on Kuramoto, and sparsity regularization gives a further gain. We have added this ablation study to the revised paper.

W4/Q2:

Main results in Table 1 use the same solver configuration across the four systems and both network sizes: the adaptive dopri5 solver with rtol = atol = 1e-3, integrating over one observation interval. Since dopri5 is adaptive, there is no fixed step size. RK4 is provided in our code as an alternative fixed-step option. Retraining with rtol = atol ∈ {1e-2, 1e-4, 1e-5} leaves structure recovery unchanged: AUROC and AUPRC remain 1.00 on all four systems (ER, N=100). We have added these details to Appendix A.4.

Q3:

We measure NOGI on Spring (ER, dopri5) with adjoint and direct backpropagation. Each cell reports epoch time / peak memory / forward,backward NFE per epoch:

N, KAdjointDirect backprop
10, 11.1s / 43MiB / 168,1680.8s / 192MiB / 168,0
100, 15.4s / 1879MiB / 168,1683.5s / 15954MiB / 168,0
100, 101.2s / 1873MiB / 52,70OOM
100, 500.8s / 449MiB / 50,1040.3s / 13616MiB / 50,0
100, 1001.3s / 232MiB / 74,2240.3s / 10030MiB / 74,0

The adjoint method uses substantially less memory than direct backpropagation. Epoch time and memory are lower at larger intervals because subsampling leaves fewer training samples. We acknowledge that the "O(1) memory cost" stated in the abstract, §1, and §2.2 was imprecise: memory stays constant as the number of solver steps (NFE) grows, but it still grows with sequence length and network size. We have corrected the memory statements in the revised version.

Thank you for the response. The complex experiment and the clarification of the Graph ablation addressed my two questions. My main concern is the methodological novelty of rough initial states for neural ODEs. For now, I will keep my original rating of 2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.