Research map / Probabilistic and scientific modeling / Causal discovery
Learning Causal Relations from Subsampled Time Series with Two Time-Slices
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 DHT-CIT, which uses two time-slices of subsampled time series as simulated interventions and a conditional HSIC independence test to recover a descendant hierarchical topology and then prune to a summary causal graph.
Problem
Causal discovery from subsampled time series with only two observed time-slices, where many unobserved time-slices lie between measurements and standard time-series or topological-ordering methods yield non-unique orderings with many spurious edges.
Approach
Treat the earlier time-slice as auxiliary instrumental variables that act as simulated interventions; a conditional independence test X_i^{ta} not-independent of X_j^{tb} given the other earlier-slice variables identifies descendants, yielding a unique descendant hierarchical topology with few spurious edges, refined by layer adjustment and CAM pruning.
Key results
- On Sin-10-10 with observational two-slice data, DHT-CIT reaches SHD 1.00±1.22 and F1 0.95±0.05, versus SCORE SHD 5.60±3.92 and CAM SHD 3.70±2.95.
- DHT-CIT reduces pruned spurious edges relative to SCORE by about 24.4 on Sin-10-10 and 147.6 on Sin-20-20.
- On the PM-CMR real dataset, only DHT-CIT recovers the dense graph consistent with prior knowledge (T<-X1:7->Y and T->Y), while GOLEM and SCORE do not.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DHT-CIT instead of GOLEM | Causal discovery on Sin-10-10 observational data · Sin-10-10 observational | SHD: 0.50±0.80 -> 1.00±1.22 | Hurt |
| Use DHT-CIT instead of GOLEM | Causal discovery on Sin-10-10 observational data · Sin-10-10 observational | F1-Score: 0.97±0.03 -> 0.95±0.05 | Hurt |
| Use DHT-CIT instead of GOLEM | Causal discovery on Sin-20-20 observational data · Sin-20-20 observational | SHD: 1.30±1.10 -> 1.00±1.32 | Helped |
| Use DHT-CIT instead of GOLEM | Causal discovery on Sin-20-20 observational data · Sin-20-20 observational | F1-Score: 0.97±0.03 -> 0.98±0.03 | No clear effect |
| Use DHT-CIT instead of CAM | Causal discovery on Sigmoid-10-10 with Gaussian noise · Sigmoid-10-10 Gaussian noise | SHD: 3.70±3.43 -> 0.67±1.12 | Helped |
| Use DHT-CIT instead of CAM | Causal discovery on Sigmoid-10-10 with Gaussian noise · Sigmoid-10-10 Gaussian noise | F1-Score: 0.82±0.17 -> 0.96±0.06 | Helped |
| Use DHT-CIT instead of CAM | Causal discovery on Poly-10-10 with Gaussian noise · Poly-10-10 Gaussian noise | SHD: 8.00±4.69 -> 3.22±3.15 | Helped |
| Use DHT-CIT instead of CAM | Causal discovery on Poly-10-10 with Gaussian noise · Poly-10-10 Gaussian noise | F1-Score: 0.63±0.21 -> 0.84±0.15 | Helped |
| Use DHT-CIT instead of GOLEM | Causal discovery on Sin-10-10 with Laplace noise · Sin-10-10 Laplace noise | SHD: 1.50±1.20 -> 1.20±1.99 | Helped |
| Use DHT-CIT instead of GOLEM | Causal discovery on Sin-10-10 with Laplace noise · Sin-10-10 Laplace noise | F1-Score: 0.92±0.05 -> 0.94±0.04 | Helped |
| Use DHT-CIT instead of ReScore | Causal discovery on Sin-10-10 with Uniform noise · Sin-10-10 Uniform noise | SHD: 1.70±0.90 -> 1.00±2.19 | Helped |
| Use DHT-CIT instead of ReScore | Causal discovery on Sin-10-10 with Uniform noise · Sin-10-10 Uniform noise | F1-Score: 0.92±0.04 -> 0.96±0.09 | Helped |
Surprising findings
- SCORE recovers near-true structure with interventional data but degrades substantially on observational two-slice data.
- Traditional time-series methods such as Granger and VARLiNGAM underperform temporal variants of non-temporal methods when only two time-slices are available.
Limitations
- Relies on acyclic summary causal graphs and stationarity (consistency throughout time) assumptions.
- Performance declines on extremely dense graphs where SCORE is recommended instead.
- Performance degrades somewhat as the subsampling rate increases.
Experimental setup
Datasets: PM-CMR
Benchmarks: PM-CMR
Metrics: SHD SID F1-Score L2 distance (Dis.) #Prune
Baselines: PC FCI GOLEM NOTEARS ReScore Granger VARLiNGAM CD-NOD CAM SCORE
Builds on: SCORE CAM NOTEARS Conditional HSIC
Closest papers by meaning
- Discovering Mixtures of Structural Causal Models from Time Series Data (ICML 2024 · similarity 0.90)
- Differentiable Causal Discovery for Latent Hierarchical Causal Models (ICLR 2025 · similarity 0.89)
- Local Identifying Causal Relations in the Presence of Latent Variables (ICML 2025 · similarity 0.88)
- Hybrid Top-Down Global Causal Discovery with Local Search for Linear and Nonlinear Additive Noise Models (NeurIPS 2024 · similarity 0.88)
- CausalRivers - Scaling up benchmarking of causal discovery for real-world time-series (ICLR 2025 · similarity 0.88)
- CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery (NeurIPS 2026 · 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.