acceptodds
Under review as a conference paper at ICLR 2027

NeSyF: Neural-Symbolic Feedback for Joint Network Dynamics Prediction and Governing Equation Discovery

Abstract

Predicting network dynamics while recovering their governing equations remains challenging because existing neural-symbolic approaches often use recovered equations as final outputs or post-hoc interpretations, leaving the interaction between prediction and equation discovery largely one-way. We propose a Neural-Symbolic Feedback(NeSyF) framework that couples network dynamics prediction and governing equation discovery. The framework decomposes network dynamics into self-dynamics and interaction dynamics, learns them with separate neural networks, and applies symbolic regression to recover explicit equations. These equations are then used as derivative-consistency supervision to refine the learned dynamics. Experiments across four network topologies and four nonlinear dynamical systems demonstrate accurate long-horizon prediction and explicit equation recovery, while experiments on two real-world epidemic datasets further demonstrate accurate state prediction and yield compact symbolic models of empirical dynamics. A rollout-error analysis provides a theoretical rationale for derivative-consistency feedback, while ablation studies demonstrate its contribution to long-horizon predictive performance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
network dynamics, graph neural networks, neural-symbolic learning, symbolic regression, governing equation discovery

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_huxnOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NeSyF, a framework combining neural networks and symbolic regression, aiming to simultaneously achieve network dynamics prediction and governance equation discovery. The method first learns the dynamics using a neural network, then recovers the symbolic equations through STLSQ and PySR, and uses these equations to optimize the neural network's predictions. Experiments covering four network topologies, four dynamical systems, and two real-world epidemic datasets demonstrate that the proposed method achieves good performance in long-term prediction and equation recovery.

Critical strengths

1.The paper explores an interesting way to use recovered symbolic equations as feedback for neural dynamics learning, rather than only for post-hoc interpretation. 2.NeSyF achieves strong long-horizon prediction performance across different network topologies and dynamical systems, with improvements over existing baselines.

Critical weaknesses
  1. While ablation experiments show that symbolic feedback improves prediction performance, its unique advantages over conventional regularization or knowledge distillation methods remain unclear. Could the authors provide a direct comparison to illustrate the additional benefits inherent in symbolic feedback itself?

  2. Appendix B provides a tighter upper bound on the error only when the condition ϵsym+ϵfb<ϵ0\epsilon_{sym}+\epsilon_{fb}<\epsilon_0 is met, but the paper does not prove or verify whether this condition holds. Could the authors verify this condition using synthetic experiments and report the corresponding errors?

  3. Table 2 only reports the results of NeSyF combined with STLSQ and PySR, and the coefficient error of NeSyF under PySR is higher than that of LaGNA (0.1745 vs. 0.0766). Could the authors supplement the results of the combined LLC and LaGNA using the same process?

  4. The actual equations in the synthetic experiment closely matched the pre-defined symbol search space, but Table 3 did not directly verify its ability to describe the dynamics of the real epidemic. Could the authors directly use the recovery equations for predictions in unseen time periods to verify whether they truly captured the underlying dynamics?

Questions for authors

See weakness

Reviewer_n8UXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper learns network dynamics in three stages. First, separate MLPs for ff and gg are trained with one-step and multi-step Euler rollout losses, giving M0M_0. Second, STLSQ and PySR are applied to M0M_0's learned functions, and an equation is chosen by Pareto selection. Third, M0M_0 is fine-tuned with a loss aligning its derivatives with those of the recovered equations, giving M1M_1; symbolic regression is then rerun on M1M_1. An appendix shows M1M_1 has a tighter worst-case Euler rollout bound if ϵsym+ϵfb<ϵ0\epsilon_{sym} + \epsilon_{fb} < \epsilon_0. Experiments use four topologies and four polynomial systems, and Japanese influenza/COVID-19 data.

Critical strengths
  • The paper is clear and the reproducibility details are thorough.
  • The ablation controls for training budget, so the rollout gain is not due to longer training.
  • Gains over LLC and LaGNA are consistent across topologies and systems.
Critical weaknesses
  • Most components seem to be existing work: the f/gf/g decomposition (Gao & Yan, LaGNA, LLC), off-the-shelf STLSQ and PySR, Pareto selection, and a derivative-matching loss. Closing the loop also has some precedent in joint neural–symbolic training (e.g., SINDy autoencoders) and in shaping networks for symbolic distillation (Cranmer et al., 2020). Combining known parts is fine in itself, but I did not find either a clear win in a harder setting or an explanation of why the combination works.
  • I am not sure the feedback loop adds new information. The equations are fit to M0M_0's outputs on the training states, M1M_1 is pulled toward them on the same states, and symbolic regression on M1M_1 then recovers roughly what it was trained toward. This would likely help when the library contains the truth, but it might produce spurious simplicity when it does not. The synthetic benchmarks may not separate these cases, since all ground-truth systems are polynomials of degree ≤3\le 3, which matches the STLSQ library.
  • The sensitivity results seem consistent with this reading. Varying rsymr_{sym} over an 8x range changes MAE by a few percent, while rsym=0r_{\text{sym}} = 0 increases it by 49–118%. I may be missing an explanation, but this looks more like a regularization effect than graded feedback.
  • As far as I can tell, equation recovery is not compared between M0M_0 and M1M_1, so it is unclear whether feedback improves equation discovery.
  • The bound follows fairly directly from the triangle inequality and a standard Euler recursion, and its condition is not checked empirically, although F∗F^* is known for the synthetic data. The condition seems close to "the symbolic fit denoises M0M_0", which is what one would want to show.
  • NeSyF leads on F1 (7 of 8 settings), but LaGNA has lower average coefficient error under both regressors (0.155 vs. 0.163 with STLSQ; 0.077 vs. 0.175 with PySR). The merged STLSQ+PySR selection appears to be applied only to NeSyF. Table 2 does not seem to state the topology or report variances.
  • Many rows of Table 1 are generic models that are orders of magnitude worse and may not be very informative comparisons. It would help to see learned f,gf, g against ground truth, error vs. horizon, and results under noise or varying data size. 200 steps at Δt=0.01\Delta t = 0.01 is about two Lyapunov times for Lorenz, so longer-horizon behavior on the attractor is not really tested. Training data are generated with explicit Euler at the same Δt\Delta t the Euler-based models use, which may favor those models. Fig. 3(a) has no error bars.
  • On real data, Rf2R^2_f and Rg2R^2_g measure agreement with the neural model rather than with the data, and M1M_1 is trained to be compressible, so these numbers may partly reflect self-consistency. On COVID-19, LLC has slightly higher R2R^2 on both. Fig. 2 shows cumulative one-step predictions, which can hide errors; quantitative forecasting metrics and simple baselines (persistence, ARIMA, metapopulation models) would make the evaluation more convincing. The recovered equations are mostly linear decay plus near-constant or linear coupling on standardized states, and I am not sure how much interpretation they support. Table 3 is hard to read.
Questions for authors
  1. What is the rollout error of the Stage-II equations used directly as a forecaster?
  2. What equation-recovery F1 and MRE does M0M_0 achieve?
  3. If Stage III is given a deliberately wrong equation, what does rerunning Stage II on M1M_1 recover?
  4. How does the feedback gain change on dynamics outside the library (e.g., Kuramoto, saturating epidemic models)?
  5. Could ϵ0\epsilon_0, ϵsym\epsilon_{sym}, ϵfb\epsilon_{fb} be estimated on the synthetic data?
  6. Why is performance nearly flat for rsym∈[0.025,0.2]r_{sym} \in [0.025, 0.2] but much worse at rsym=0r_{sym} = 0?
Reviewer_DrVnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes NeSyF, a neural-symbolic feedback framework for networked dynamical systems. Existing neural-symbolic pipelines are mostly one-way: a neural network first fits the dynamics, and symbolic regression is then used only as post-processing to extract equations, without feeding the recovered symbolic equations back to improve the network. NeSyF instead treats the recovered symbolic governing equations as supervisory signals and feeds them back into the neural network, forming a bidirectional closed loop.

Critical strengths

A novel bidirectional neural-symbolic closed-loop paradigm. Existing neural-symbolic methods for networked dynamics are one-way: from the neural network to symbolic equations (only as output). NeSyF innovatively uses the extracted symbolic equations to construct a derivative consistency loss, which back-supervises neural network optimization to form a closed loop; it also provides theoretical support for rollout error propagation under Euler integration, explaining why symbolic feedback benefits long-horizon prediction.

Critical weaknesses

1.The performance ceiling of symbolic feedback depends on the quality of symbolic regression in Stage II. Stage III training completely fixes the symbolic equations output by Stage II. If the equations obtained by symbolic regression in Stage II are themselves of poor quality or contain systematic bias, L_sym will align the neural network toward an incorrect vector field, degrading prediction performance. The paper lacks an analysis of this failure mode and corresponding experiments, and does not investigate the performance degradation of NeSyF when symbolic regression quality is low.

2.Symbolic regression is a post-processing step, and its search space and operator library must be manually predefined, which constrains generality. Both ST-LSQ and PySR in Stage II rely on manually specified function libraries. For completely unknown dynamics, the setting of the operator library greatly affects the quality of the recovered equations; the paper does not investigate the robustness of the method when the operator library does not match the true dynamics. In addition, the framework is built on the strong prior that dynamics can be decomposed into self-dynamics plus pairwise interactions, and cannot natively handle higher-order interactions. Moreover, the paper does not analyze the reasons for choosing ST-LSQ and PySR; the claim that the two paradigms are merely complementary does not seem very convincing. There are many types of existing symbolic regression methods, such as Transformer-based and LLM-based methods, and an ablation study on the choice of methods is needed.

3.There is a lack of robustness experiments on noise, topological perturbations, and similar factors. In all experiments, the network topology is completely known as ground truth. In reality, networked dynamics data commonly exhibit node state observation noise and topological noise (missing edges and spurious edges). The paper does not evaluate under these two types of noise perturbations, making it difficult to fully demonstrate the robustness of the method.

4.The comparison baselines lack related work such as ND^2 and SymBOL, which are related symbolic regression works for network dynamics.

Questions for authors

1.The Stage-III training process fixes the symbolic equations obtained in Stage-II throughout. When the symbolic regression output of Stage-II is of poor quality or contains systematic errors, the derivative consistency loss L_sym may mislead the updating of the neural network. We hope the authors can add relevant experiments.

2.ST-LSQ and PySR in Stage-II rely on a manually specified operator set. If the true dynamics contain functions that are not in the operator library, this will seriously affect the Stage-II output. Can the authors add ablation experiments under different operator libraries, and verify and analyze the performance when the operator library does not match the true dynamical equations?

3.All current experiments use noise-free observations and fully known ground-truth topology. On synthetic datasets, please add observation noise experiments: add noise of different intensities and different types to node states, and also add topology perturbation noise experiments, evaluating the performance of the proposed method by adding topological noise.

4.The overall pipeline in this paper is a single one-way execution process. After Stage-III completes neural network optimization, symbolic regression is rerun only once, and no multi-round iterative closed loop is constructed; that is, the updated M1 is not used to run Stage-II again to obtain new symbolic equations and then feed them back to optimize the network again. Is it possible to perform multi-round iterative feedback and conduct relevant analysis?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.