Research map / Probabilistic and scientific modeling / Causal discovery
Discovering Latent Causal Graphs from Spatiotemporal 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.
Proposes SPACY, a variational-inference causal representation learning framework that jointly infers latent time series, spatial RBF factors, and a latent causal DAG from gridded spatiotemporal data, with identifiability results on continuous spatial domains.
Problem
Causal discovery from high-dimensional spatiotemporal gridded time series is hard due to scalability and spatial correlations among nearby grid points, which hurt conditional independence tests; existing two-stage or single-parent latent methods are limited.
Approach
Model observed gridded series as a grid-pointwise nonlinear mixing of latent time series via RBF spatial factors, learn latent SCM (Rhino-based) and spatial factors end-to-end via ELBO, and prove identifiability on a continuous spatial domain allowing instantaneous edges and multiple latent parents.
Key results
- SPACY outperforms baselines (Mapped-PCMCI, Linear-Response, LEAP, TDRL, CDSD) in orientation F1 across D=10,20,30 on synthetic data, with largest gaps in nonlinear settings.
- SPACY scales to grids up to 250x250 with moderate runtime growth, while Linear-Response scales poorly.
- On climate data, SPACY recovers subgraphs consistent with MJO, NAO, and AAO teleconnections qualitatively; no ground-truth evaluation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SPACY instead of the evaluated baselines | Synthetic data with linear and nonlinear SCMs · Synthetic data | orientation F1 | Helped |
| Use SPACY instead of CDSD or Mapped PCMCI | Synthetic data · Synthetic data | MCC | Mixed |
| Use SPACY instead of LEAP, TDRL, or Linear-Response | Synthetic data · Synthetic data | MCC | Helped |
| Use SPACY instead of LEAP, TDRL, or Linear-Response as latent dimension increases | Synthetic data · Synthetic data | orientation F1 | Helped |
| Use SPACY instead of Mapped PCMCI | Causal discovery on increasing grid sizes · Synthetic data, grid-size scalability | orientation F1 | Helped |
| Use SPACY instead of LEAP, TDRL, or CDSD | Causal discovery on increasing grid sizes · Synthetic data, grid-size scalability | orientation F1 | Helped |
| Use SPACY instead of Linear-Response | Causal discovery on increasing grid sizes · Synthetic data, grid-size scalability | runtime | Helped |
Surprising findings
- Linear-Response and LEAP/TDRL degrade sharply in nonlinear settings as latent dimension grows.
Limitations
- Identifiability theory assumes a continuous spatial domain; finite-grid guarantees are heuristic.
- Isotropic RBF kernel is a modeling simplification for localized climate processes.
- Evaluation on real climate data is qualitative without ground-truth causal graphs.
Experimental setup
Datasets: synthetic spatiotemporal datasets (linear and nonlinear SCM, D=10/20/30) Global Climate Dataset
Benchmarks: synthetic spatiotemporal datasets Global Climate Dataset
Metrics: orientation F1 mean correlation coefficient (MCC) runtime
Baselines: Mapped-PCMCI Linear-Response LEAP TDRL CDSD
Builds on: Rhino LEAP PCMCI NOTEARS
Closest papers by meaning
- SPACETIME: Causal Discovery from Non-Stationary Time Series (AAAI 2025 · similarity 0.91)
- Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model (NeurIPS 2024 · similarity 0.89)
- Discovering Mixtures of Structural Causal Models from Time Series Data (ICML 2024 · similarity 0.88)
- CausalRivers - Scaling up benchmarking of causal discovery for real-world time-series (ICLR 2025 · similarity 0.88)
- Local Identifying Causal Relations in the Presence of Latent Variables (ICML 2025 · similarity 0.88)
- Identifying Macro Conditional Independencies and Macro Total Effects in Summary Causal Graphs with Latent Confounding (AAAI 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.