Physics-Informed High-order Graph Dynamics Identification Learning for Predicting Complex Networks Long-term Dynamics

Basic Information

• Paper ID: 2510.09082
• Title: Physics-Informed High-order Graph Dynamics Identification Learning for Predicting Complex Networks Long-term Dynamics
• Authors: Bicheng Wang, Junping Wang, Yibo Xue (Institute of Automation, Chinese Academy of Sciences)
• Classification: cs.AI cs.CY cs.SI physics.soc-ph
• Publication Date: October 2025 (ArXiv Preprint)
• Paper Link: https://arxiv.org/abs/2510.09082

Abstract

This paper proposes a Physics-informed High-order graph Structure Learning method (PhyHSL) for predicting long-term dynamic evolution of complex networks. The method captures high-order non-pairwise relationships in networks through dynamic hypergraph learning and constructs a dual-driven prediction module by combining Koopman operator theory and physics-informed neural differential equations, enhancing model interpretability while ensuring prediction accuracy. The effectiveness of the method is validated through experiments on public datasets and self-constructed industrial supply chain network datasets.

Research Background and Motivation

Core Problems

Complex network dynamics prediction faces two major challenges:

1. Limitations in Relationship Modeling: Existing methods typically use simple graphs to describe network relationships, capturing only pairwise interactions, while complex networks contain abundant non-pairwise structural relationships (e.g., multi-enterprise cooperation in supply chains, network structures in transportation networks).
2. Trade-off Between Prediction Accuracy and Interpretability: Theoretical prediction models lack accuracy, data-driven models lack interpretability, and long-term predictions are prone to error accumulation.

Research Significance

Complex network dynamics learning is crucial for understanding, modeling, and controlling real-world complex systems, involving brain networks, social networks, supply networks, and other domains. Accurate prediction of network evolution helps analyze network resilience and forecast future states.

Limitations of Existing Methods

• First-order GNN Constraints: Traditional graph neural networks struggle to capture dynamic non-pairwise relationships
• Hypergraph Method Dependencies: Existing hypergraph methods largely rely on predefined structures and cannot adapt to dynamic network evolution characteristics
• Missing Physical Constraints: Pure data-driven methods lack physical mechanism constraints, and predictions easily deviate from actual system evolution trajectories

Core Contributions

1. Dynamic Hypergraph Structure Learning Module: Breaks through the limitation of traditional hypergraphs relying on predefined structures by dynamically generating adaptive hyperedges through low-rank matrix factorization and hypergraph convolution, enabling online modeling of non-pairwise interactions.
2. Physics-Data Dual-Driven Prediction Module:
   • Introduces Koopman operator theory to convert nonlinear dynamic differential equations into linear system solutions
   • Utilizes physics-informed neural ODEs to ensure dynamic evolution complies with physical laws
   • Achieves collaborative optimization through variational inference framework, enhancing model robustness
3. Complete High-order Network Dynamics Framework: Integrates physical laws with data-driven high-order graph structure modeling, constructing a collaborative optimization paradigm for long-term complex network dynamics prediction.
4. Experimental Validation: Validates prediction accuracy and generalization performance on public datasets and self-constructed supply chain network datasets.

Method Details

Task Definition

Given historical observation sequences of a complex network, predict the network's future dynamic evolution states. Input consists of node feature sequences and network structure, with output being node state predictions at future time steps.

Model Architecture

1. Domain Relationship Capture Encoder

Constructs temporal graphs containing spatial and temporal edges:

• Spatial Edges: Weighted edges between objects at the same timestamp
• Temporal Edges: Weighted edges between consecutive observations of each object

Adjacency matrix defined as:

A(it, jt') = {
    w^t_ij,  if t' = t
    1,       if i = j, t' = t+1  
    0,       otherwise
}

First-order Spatial Convolution: Combines attention mechanisms to adaptively learn neighborhood information

s^(k)(it, jt') = A(it, jt') cos(W_query h^{t,(k)}_i, W_key h^{t',(k)}_j)
h^{t,(k+1)}_i = h^{t,(k)}_i + σ(∑_{jt'∈N_it} s^(k)(it, jt') W_value h^{t',(k)}_j)

Second-order Spectral Convolution: Uses Chebyshev polynomials to explore non-neighborhood semantic information

C^(k) = ∑^2_{m=0} T_m(L̃)C^{(k-1)}W^(k)_m

2. Dynamic Hypergraph Structure Learning (DHSL)

Dynamically generates hyperedges through low-rank matrix factorization:

Λ = UW_Λ

where U is the stacked node state representation and W is a learnable weight matrix.

Hypergraph convolution process:

E = σ(W_E Λ^T U) + Λ^T U  (hyperedge embedding)
F_i = ΛE = Λ(σ(W_E Λ^T U) + Λ^T U)  (node embedding update)

3. Network Dynamics Learning

Physics-Informed Driven Module:

• Generates initial states through variational inference: q(z^0_i|X,A) = N(MLP_m(f_i), MLP_v(f_i))
• Computes future states using neural ODE solver: (z^1_i, z^2_i, ..., z^{T+1}_i) = ODESolver(z^0_i, g, [t=0,...,T])

Data-Driven Module: Based on Koopman operator theory, maps nonlinear systems to linear space:

K ∘ g(x_t) = g(F(x_t)) = g(x_{t+1})
(z̃^0_i, z̃^1_i, ..., z̃^{T+1}_i) = (f^0_i, Kf^0_i, Kf^1_i, ..., Kf^T_i)

Fused Prediction:

x̂^t_i = MLP(σ([z^t_i, z̃^t_i]))

Technical Innovations

1. Dynamic Hypergraph Generation: No predefined structures required; hyperedges are adaptively generated from node states
2. Physical Constraints Combined with Linearization: Koopman operator provides global linearized representation; neural ODE ensures physical consistency
3. Collaborative Optimization Framework: Dual-driven modules are jointly trained under variational inference framework

Experimental Setup

Datasets

Public Datasets:

• Social (Facebook pages): 3,892 nodes, 17,239 edges
• Web (EPA links): 4,252 nodes, 8,896 edges
• WS (Watts-Strogatz network): 5,000 nodes, 10,000 edges

Self-Constructed Supply Chain Datasets:

• Manufacture: 960 nodes, 25,142 edges
• Electronic: 700 nodes, 16,604 edges
• Finance: 1,500 nodes, 61,218 edges

Evaluation Metrics

Uses Mean Absolute Error (MAE):

MAE = (1/N) ∑^N_{i=1} ||x̂_i - x_i||

Baseline Methods

• GNN Methods: DCRNN, MTGODE, DiskNet
• Hypergraph Methods: HGC-RNN, MSHyper
• PINN Methods: PhyCRNet, PINNsFormer, PhysicsSolver

Implementation Details

• Framework: PyTorch
• Hardware: 2 NVIDIA A100 GPUs
• Repeated experiments 10 times for averaged results
• Optimization: Variational inference framework, minimizing ELBO loss

Experimental Results

Main Results

PhyHSL achieves optimal or near-optimal results on all 6 datasets:

Public Dataset Performance:

• Social: 0.201±0.007 (optimal)
• Web: 0.178±0.014 (optimal)
• WS: 0.127±0.007 (optimal)

Supply Chain Dataset Performance:

• Manufacture: 0.112±0.014 (optimal)
• Electronic: 0.247±0.013 (optimal)
• Finance: 0.162±0.027 (near-optimal)

Achieves approximately 10% average improvement over best baseline methods, with more pronounced advantages on complex supply chain networks.

Ablation Studies

Ablation studies on Social and Manufacture datasets show:

• Removing physics-driven module: performance drops (0.231 vs 0.201)
• Removing Koopman module: performance drops (0.233 vs 0.201)
• Removing hypergraph module: more significant impact on complex networks
• Removing both dual-driven modules: substantial performance drop (0.268 vs 0.201)

Long-term Prediction Analysis

• Training Length Impact: Prediction performance improves and stabilizes with increasing training length
• Prediction Length Impact: PhyHSL shows more pronounced advantages over DiskNet in long-term predictions
• Computational Efficiency: Higher computational efficiency compared to Transformer-dependent methods

Experimental Findings

1. Dual-driven modules complement each other; both are essential
2. Hypergraph module plays a more prominent role in complex networks
3. Physical constraints effectively reduce error accumulation in long-term predictions
4. Koopman operator reduces learnable parameters and improves computational efficiency

Related Work

Network Dynamics Prediction

• Early methods based on first-order GNNs, such as NCDN first combining neural ODEs and GNNs
• MTGODE abstracts multivariate time series as dynamic graphs
• DiskNet identifies network skeletons based on renormalization group structures in hyperbolic space

Hypergraph Neural Networks

• HGNN: first spatial hypergraph learning method
• DHGNN: first to handle dynamic hyperedges
• Existing methods largely rely on predefined structures or node similarity

Conclusions and Discussion

Main Conclusions

1. PhyHSL effectively integrates physical constraints and high-order structure learning, significantly improving long-term complex network dynamics prediction performance
2. Dynamic hypergraph learning successfully captures non-pairwise relationships; dual-driven modules ensure accuracy and interpretability
3. Demonstrates good practical value in industrial scenarios

Limitations

1. High model complexity requiring balance between performance and computational cost
2. Applicability to extremely sparse or ultra-large-scale networks remains to be verified
3. Physical constraint design may require domain expert knowledge

Future Directions

1. Explore more complex hypergraph network relationship construction
2. Investigate online learning methods for real-time network structure updates
3. Develop real-time monitoring and regulation technologies for network resilience

In-Depth Evaluation

Strengths

1. Strong Method Innovation: First to organically combine Koopman operator, physics-informed neural ODEs, and dynamic hypergraph learning
2. Clear Problem Definition: Accurately identifies core challenges in complex network prediction
3. Comprehensive Experimental Design: Covers public and self-constructed datasets with sufficient ablation studies
4. Reasonable Technical Approach: Integration of physical constraints and data-driven methods has theoretical foundation

Weaknesses

1. Insufficient Theoretical Analysis: Lacks theoretical guarantees for convergence and stability
2. Missing Computational Complexity Analysis: No detailed complexity analysis provided
3. Hyperparameter Sensitivity: Insufficient discussion of critical hyperparameter impacts
4. Limited Interpretability Verification: Physical constraint effectiveness validation is insufficient

Impact

1. Academic Contribution: Provides new technical paradigm for complex network dynamics prediction
2. Practical Value: Demonstrates application potential in industrial scenarios such as supply chains
3. Reproducibility: Provides detailed implementation details facilitating reproduction

Applicable Scenarios

• Supply chain network risk prediction and management
• Social network information propagation modeling
• Traffic network flow prediction
• Biological network dynamics analysis
• Financial network risk propagation

References

The paper cites important works in related fields, including:

• Graph neural network foundational methods (Kipf et al., Veličković et al.)
• Neural ODE theory (Chen et al.)
• Koopman operator theory (Mezić, Strogatz)
• Hypergraph neural networks (Feng et al., Jiang et al.)
• Physics-informed neural networks (Raissi)

Overall Assessment: This paper proposes a technically innovative and practically valuable complex network dynamics prediction framework with comprehensive method design and experimental validation. Despite limitations in theoretical analysis and computational complexity discussion, its technical contributions and application prospects remain noteworthy.