Research map / Probabilistic and scientific modeling / Dynamical systems
Benchmarking Structural Inference Methods for Interacting Dynamical Systems with Synthetic Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Benchmarks 13 structural inference methods (statistics, information theory, tree-based, deep learning) on a new synthetic dataset DoSI (Springs, NetSims, Charged Particles dynamics over 11 graph types) plus a real EMT scRNA-seq dataset, measuring AUROC for accuracy, scalability, and robustness to noise.
Problem
Existing structural inference methods for interacting dynamical systems are domain-specific, evaluated on distinct datasets, and lack a standardized, objective comparison framework.
Approach
Introduce the Dataset for Structural Inference (DoSI), with trajectories generated over 11 types of real-world-inspired interaction graphs and multiple simulators, and use it to evaluate 13 methods from different disciplines under one unified AUROC-based protocol.
Key results
- Deep learning methods (NRI, ACD, MPM, iSIDG, RCSI) outperform on multi-dimensional Springs/NetSims data, e.g., NRI shows 46.35% higher AUROC on gene coexpression networks on the multi-dimensional setting.
- ppcor and TIGRESS show stable performance across graph types and sizes, with ppcor dropping only 0.71% from 15 to 100 nodes.
- Deep learning methods lose at least 7.94% AUROC from 15 to 100 nodes, while information-theory methods like PIDC and Scribe show robustness to Gaussian noise.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use multi-dimensional trajectories rather than one-dimensional trajectories | NRI · GCN | AUROC: 46.35% higher AUROC | Helped |
| Use multi-dimensional trajectories rather than one-dimensional trajectories | ACD · LN | AUROC: 34.30% increase | Helped |
| Increase graph size from 15 to 100 nodes | Deep learning methods · DoSI, Springs | AUROC: 7.94% | Hurt |
| Increase graph size from 15 to 100 nodes | ppcor · DoSI, graph-size scalability | AUROC: 0.71% | Hurt |
| Add Gaussian noise to trajectories | TIGRESS, CLR, and PIDC · BN trajectories | AUROC | No clear effect |
| Add Gaussian noise to trajectories | Deep learning methods · BN trajectories | AUROC | Hurt |
| Use trajectories from sparser graphs with longer average shortest path length | Structural inference methods · DoSI interaction graphs | AUROC | Helped |
| Infer two distinct edge types rather than a single edge type | Deep learning methods · Charged Particles | AUROC | Hurt |
Surprising findings
- PIDC and dynGENIE3 improve on larger graphs, contrary to the general degradation trend.
- Performance correlates positively with average shortest path length and negatively with average degree.
Limitations
- Relies on static graph assumptions; dynamic topologies not covered.
- Deep learning methods not adapted to multi-dimensional Charged Particles data for non-DL methods; only one-dimensional data for statistical, information-theory and tree methods.
- Subsets with more than 100 nodes are not publicly downloadable; only partial dataset released.
Experimental setup
Datasets: DoSI (synthetic Springs, NetSims, Charged Particles trajectories) EMT scRNA-seq dataset (Sha et al. / Cook and Vanderhyden) STRING database ground-truth GRN
Benchmarks: DoSI EMT dataset
Metrics: AUROC delta AUROC under noise
Baselines: ppcor TIGRESS ARACNe CLR PIDC Scribe dynGENIE3 XGBGRN
Builds on: NRI dynGENIE3 GENIE3 ARACNe CLR
Closest papers by meaning
- A Graph Dynamics Prior for Relational Inference (AAAI 2024 · similarity 0.89)
- Structural Inference of Dynamical Systems with Conjoined State Space Models (NeurIPS 2024 · similarity 0.89)
- Interpretable Relational Inference with LLM-Guided Symbolic Dynamics Modeling (NeurIPS 2026 · similarity 0.88)
- Guided Structural Inference: Leveraging Priors with Soft Gating Mechanisms (ICML 2025 · similarity 0.87)
- Let's do the time-warp-attend: Learning topological invariants of dynamical systems (ICLR 2024 · similarity 0.86)
- CausalDynamics: A large‐scale benchmark for structural discovery of dynamical causal models (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.