Research map / Probabilistic and scientific modeling / Causal discovery
SPACETIME: Causal Discovery from Non-Stationary Time Series
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.
Proposes SpaceTime, an MDL-based score-and-search method using non-parametric GP functional models and kernelized HSIC tests to jointly discover temporal causal graphs, regime changepoints, and context partitions from non-stationary multi-dataset time series.
Problem
Causal graph discovery from non-stationary time series collected across multiple contexts (locations), where causal mechanisms change over time (regimes) and across datasets, without assuming stationarity or known changepoints.
Approach
Unifies causal discovery, changepoint detection and partitioning of datasets/time into contexts and regimes under a single MDL score built from GP mechanism descriptions, with changepoints found from GP prediction-error changes (PELT) and partitions from kernelized HSIC tests, iterated to convergence.
Key results
- On synthetic data, SpaceTime consistently performs well against baselines on changepoint F1, edge F1 and regime partition ARI/NMI in the non-linear Gaussian setting.
- On river discharge, SpaceTime discovers a lagged edge P(t-1) to Q(t) common across 307 European catchments, with regional variation in causal strength.
- On FLUXNET, the method recovers a 2003 European heatwave regime at Hainich and Tharandt, and the t-SNE of edge strengths tracks radiation and precipitation rather than geography.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SpaceTime instead of R-PCMCI for changepoint detection and regime partitioning | Synthetic time-series settings · Synthetic data | F1 score for changepoint detection; ARI and NMI for regime partitioning | Helped |
| Use J-PCMCI+ to model multiple contexts instead of single-context methods | Synthetic causal discovery · Synthetic data | F1 score for causal discovery | No clear effect |
| Estimate the causal graph rather than provide the true graph to SpaceTime | SpaceTime · Synthetic changepoint detection and regime partitioning | F1 score for changepoint detection; ARI and NMI for regime partitioning | No clear effect |
Surprising findings
- Grouping biosphere-atmosphere interactions by causal strength correlates with water availability and radiation rather than geographic location or ecosystem type.
Limitations
- Relies on assumptions such as sufficient sample size per regime/context, minimal regime duration, and no unobserved confounding.
- Evaluated on synthetic data with linear or non-linear Gaussian and uniform noise settings; no ground truth available for real-world case studies.
Experimental setup
Datasets: GRDC river discharge data (307 European catchments) FLUXNET biosphere-atmosphere flux data (63-64 locations) synthetic data generated with Tigramite
Benchmarks: synthetic non-stationary time series GRDC river discharge FLUXNET
Metrics: F1 ARI NMI
Baselines: GP-KNDAG R-PCMCI PCMCI+ VARLiNGAM DYNOTEARS CD-NOD J-PCMCI+
Builds on: PCMCI+ R-PCMCI J-PCMCI+ CD-NOD PELT
Closest papers by meaning
- Discovering Latent Causal Graphs from Spatiotemporal Data (ICML 2025 · similarity 0.91)
- CausalRivers - Scaling up benchmarking of causal discovery for real-world time-series (ICLR 2025 · similarity 0.90)
- Causal discovery with endogenous context variables (NeurIPS 2024 · similarity 0.90)
- CausalDynamics: A large‐scale benchmark for structural discovery of dynamical causal models (NeurIPS 2025 · similarity 0.89)
- Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model (NeurIPS 2024 · similarity 0.88)
- Flow based approach for Dynamic Temporal Causal models with non-Gaussian or Heteroscedastic Noises (NeurIPS 2025 · similarity 0.88)
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.