Research map / Probabilistic and scientific modeling / Causal discovery
CUTS+: High-Dimensional Causal Discovery from Irregular 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.
Extends the Granger-causality causal discovery method CUTS to high-dimensional irregularly sampled time series via coarse-to-fine group discovery (C2FD) and a message-passing GNN predictor (MPGNN) that shares parameters across targets.
Problem
Time-series causal discovery degrades on high-dimensional data (N > 20) and with missing entries, because neural Granger-causality models are redundant (N separate networks) and the causal graph becomes huge.
Approach
Start with coarse groups of time series and progressively split them (C2FD) to shrink the learnable causal graph, and replace component-wise MLP/LSTM predictors with a parameter-shared message-passing GRU (MPGNN) that uses the sampled binary causal graph as adjacency, while jointly imputing missing values.
Key results
- On VAR with N=128 and no missing values, CUTS+ reaches AUROC 0.9972 vs 0.9626 for CUTS and 0.9168 for NGC (ZOH).
- On Lorenz-96 with N=256 under RM p=0.6, CUTS+ achieves 0.9916 AUROC vs 0.8668 for CUTS.
- On Dream-3 (N=100, no missing), CUTS+ achieves AUROC 0.6374 vs 0.5915 for CUTS and 0.5579 for NGC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace CUTS prediction network with MPGNN (CUTS with C2FD to CUTS+) (ablation) | CUTS with C2FD · VAR, RM p=0.3 | AUROC: 0.9594 -> 0.9907 | Helped |
| replace CUTS prediction network with MPGNN (CUTS with C2FD to CUTS+) (ablation) | CUTS with C2FD · VAR, RM p=0.6 | AUROC: 0.8752 -> 0.9569 | Helped |
| replace CUTS prediction network with MPGNN (CUTS with C2FD to CUTS+) (ablation) | CUTS with C2FD · VAR, RBM p_blk=0.15% | AUROC: 0.9742 -> 0.9939 | Helped |
| replace CUTS prediction network with MPGNN (CUTS with C2FD to CUTS+) (ablation) | CUTS with C2FD · VAR, RBM p_blk=0.3% | AUROC: 0.9651 -> 0.9912 | Helped |
| replace CUTS prediction network with MPGNN (CUTS with C2FD to CUTS+) (ablation) | CUTS with C2FD · VAR, no missing | AUROC: 0.9875 -> 0.9972 | Helped |
| add C2FD to CUTS (ablation) | CUTS · VAR, RM p=0.3 | AUROC: 0.9434 -> 0.9594 | Helped |
| add C2FD to CUTS (ablation) | CUTS · VAR, RM p=0.6 | AUROC: 0.8814 -> 0.8752 | Hurt |
| add C2FD to CUTS (ablation) | CUTS · VAR, RBM p_blk=0.15% | AUROC: 0.9579 -> 0.9742 | Helped |
| add C2FD to CUTS (ablation) | CUTS · VAR, RBM p_blk=0.3% | AUROC: 0.9505 -> 0.9651 | Helped |
| add C2FD to CUTS (ablation) | CUTS · VAR, no missing | AUROC: 0.9626 -> 0.9875 | Helped |
| add C2FD to CUTS (ablation) | CUTS · Lorenz-96, RM p=0.3 | AUROC: 0.9360 -> 0.9790 | Helped |
| add C2FD to CUTS (ablation) | CUTS · Lorenz-96, RM p=0.6 | AUROC: 0.8668 -> 0.9069 | Helped |
Surprising findings
- Adding C2FD alone to CUTS sometimes lowers AUROC (e.g., VAR RM p=0.6: 0.8752 vs 0.8814 for CUTS).
Limitations
- Assumes no latent confounders and no instantaneous effects.
- Performance on real AQI data cannot be quantitatively verified without ground truth.
- Dream-3 time series are too short (T=21) for missing-data experiments.
Experimental setup
Datasets: VAR simulated Lorenz-96 simulated Dream-3 Air Quality (AQI)
Benchmarks: VAR (N=128) Lorenz-96 (N=256) Dream-3 (N=100) Air Quality (qualitative)
Metrics: AUROC
Baselines: NGC eSRU PCMCI LCCM NGM SCGL CUTS TimesNet imputation zero-order hold imputation
Builds on: CUTS cMLP / cLSTM (neural Granger causality) Gumbel-Softmax GRU message-passing neural network
Closest papers by meaning
- Towards the Disappearing Truth: Fine-Grained Joint Causal Influences Learning with Hidden Variable-Driven Causal Hypergraphs in Time Series (AAAI 2024 · similarity 0.89)
- CausalTime: Realistically Generated Time-series for Benchmarking of Causal Discovery (ICLR 2024 · similarity 0.89)
- Differentiable Causal Discovery for Latent Hierarchical Causal Models (ICLR 2025 · similarity 0.87)
- Learning Causal Relations from Subsampled Time Series with Two Time-Slices (ICML 2024 · similarity 0.87)
- CausalRivers - Scaling up benchmarking of causal discovery for real-world time-series (ICLR 2025 · similarity 0.87)
- Discovering Mixtures of Structural Causal Models from Time Series Data (ICML 2024 · similarity 0.87)
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.