Research map / Probabilistic and scientific modeling / Causal discovery
Influence without Confounding: Causal Discovery from Temporal Data with Long-term Carry-over Effects
Fan Li, Zixuan Liu, Yi Zhao, Qi Tan, Jinyang Li, Ke Xu, Shuihai Hu, jingbin zhou, TAN Kun
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LEVER (RL ordering on R matrix) versus best static baseline on G(10,20) | LEVER · G(10,20) static | F1: 0.8179 -> 0.8899 (+8.79%) | Improved |
| LEVER versus best baseline SHD on G(10,20) | LEVER · G(10,20) static | SHD: 6.67 -> 4.00 (-40.00%) | Improved |
| LEVER versus best baseline on G(20,40) | LEVER · G(20,40) static | F1: 0.7918 -> 0.8337 (+5.30%) | Improved |
| LEVER versus best baseline FPR on G(20,40) | LEVER · G(20,40) static | FPR: 0.0062 -> 0.0090 | Hurt |
| use historical refinement on temporal data versus best baseline | LEVER · Temporal synthetic | SHD and F1: 64%-88% SHD reduction, 45%-101% F1 improvement | Improved |
| use R matrix state instead of neural encoder (efficiency) | LEVER vs CORL · G(20,40) static, 20 nodes | time per episode (s): 1.71 -> 0.25 | Improved |
| use R matrix state instead of neural encoder (memory) | LEVER vs CORL · G(20,40) static, 20 nodes | peak memory (MB): 64.68 -> 10.36 | Improved |
| refine with history instead of unfolding lagged variables | LEVER · 10-node temporal graphs | extra runtime: 0.0017 seconds extra versus 203x and 69x for NOTEARS and PC | Improved |
| replace RL variable reordering with PC, NOTEARS, RL, CORL or SCORE (LV variants) (ablation) | LEVER · Temporal synthetic (linear and exponential decay) | F1 and accuracy | Improved |
| violate recurrence assumption by zeroing dependencies except at lag 25 (ablation) | LEVER · Temporal synthetic | SHD: 4 -> 6 | No clear effect |
| violate assumption except at lag 10 (ablation) | LEVER · Temporal synthetic | F1: 0.8947 -> 0.9231 | No clear effect |
| increase window size up to h+1 (ablation) | LEVER · Complicated-decay temporal data (h=10, 20) | ROC accuracy | Improved |
Closest papers by meaning
- CausalRivers - Scaling up benchmarking of causal discovery for real-world time-series (ICLR 2025 · similarity 0.90)
- On the Recoverability of Causal Relations from Temporally Aggregated I.I.D. Data (ICML 2024 · similarity 0.89)
- Discovering Mixtures of Structural Causal Models from Time Series Data (ICML 2024 · similarity 0.89)
- UnCLe: Towards Scalable Dynamic Causal Discovery in Non-linear Temporal Systems (NeurIPS 2025 · similarity 0.89)
- CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery (NeurIPS 2026 · similarity 0.89)
- CausalDynamics: A large‐scale benchmark for structural discovery of dynamical causal models (NeurIPS 2025 · 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.
