Research map / Probabilistic and scientific modeling / Causal discovery
CausalDynamics: A large‐scale benchmark for structural discovery of dynamical causal models
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.
Introduces CausalDynamics, a tiered benchmark and generator of thousands of coupled ODE/SDE and climate-model causal graphs, and benchmarks ten causal discovery methods on graph reconstruction.
Problem
Causal discovery for nonlinear, noisy, confounded, lagged dynamical systems lacks standardized benchmarks with known ground-truth causal graphs; existing benchmarks are static, low-dimensional, or weakly nonlinear.
Approach
A tiered framework that builds ground-truth dynamical causal graphs from 3D chaotic ODE/SDE systems (dysts), couples them via scale-free DAGs (Growing Network with Redirection) with MLP edges, periodic/linear drivers, confounders, time-lags and standardization, plus two idealized climate models (XRO, qgs/MAOOAM).
Key results
- On the coupled default setting, PCMCI+ and F-PCMCI reach AUROC .67 and AUPRC .25/.27, and CUTS+ attains the lowest SHD of 152.00 among methods in the coupled standardize/default comparison (lowest SHD on coupled default).
- Non-deep-learning methods such as PCMCI+ achieve the best AUROC on climate MAOOAM (.69 / .88), outperforming DL-based methods.
- Many baselines drop to near-chance AUROC (~0.50) on coupled and climate experiments with noise, confounding, or time-lags, and the benchmark contains over 14,000 graphs (14,693 total).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use simpler non-DL causal discovery algorithms rather than advanced DL-based algorithms | Causal discovery on CausalDynamics · CausalDynamics benchmark | graph reconstruction performance | Helped |
| Standardize the generated time series | Coupled causal discovery experiments · Coupled systems | AUROC | No clear effect |
| Use topology-based TSCI rather than purely DL-based approaches | Causal discovery algorithms · Confounded and higher-dimensional systems | graph reconstruction performance | Helped |
Surprising findings
- Simpler non-DL methods often outperform deep-learning-based causal discovery methods, especially on high-dimensional climate data.
- Minimal improvement is observed for standardization, even for methods like VARLiNGAM that exploit topological ordering.
Limitations
- Only a fixed set of parameters (time-lag, noise level) and 3-dimensional systems were processed for the released data.
- AUROC can be deceptively high for sparse graphs due to many true negatives, so AUPRC is recommended; SHD is not normalized across graph sizes.
Experimental setup
Datasets: dysts ODE/SDE systems XRO ENSO model output qgs MAOOAM model output ORAS5 reanalysis (for XRO parameterization)
Benchmarks: CausalDynamics simple tier CausalDynamics coupled tier CausalDynamics climate tier
Metrics: AUROC AUPRC SHD
Baselines: PCMCI+ F-PCMCI VARLiNGAM DYNOTEARS NGC TSCI CUTS+ RCD GRaSP TCDF
Builds on: dysts Growing Network with Redirection XRO MAOOAM PCMCI+
Closest papers by meaning
- Discovering Mixtures of Structural Causal Models from Time Series Data (ICML 2024 · similarity 0.89)
- SPACETIME: Causal Discovery from Non-Stationary Time Series (AAAI 2025 · similarity 0.89)
- Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed Time Series (ICML 2025 · similarity 0.89)
- CausalTime: Realistically Generated Time-series for Benchmarking of Causal Discovery (ICLR 2024 · similarity 0.89)
- Tangent Space Causal Inference: Leveraging Vector Fields for Causal Discovery in Dynamical Systems (NeurIPS 2024 · similarity 0.89)
- CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery (NeurIPS 2026 · similarity 0.89)
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.