Research map / Graph learning / Graph algorithms and clustering
Learning to Induce Causal Structure
Nan Rosemary Ke, Silvia Chiappa, Jane X Wang, Jorg Bornschein, Anirudh Goyal, Melanie Rey, Theophane Weber, Matthew Botvinick, Michael Curtis Mozer, Danilo Jimenez Rezende
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| train CSIvA on synthetic Dirichlet data and apply it to Sachs | CSIvA · Sachs | Hamming distance: 4 vs 13 | Improved |
| train CSIvA on synthetic MLP data and apply it to Asia | CSIvA · Asia | Hamming distance: 3 vs 4 | Improved |
| use CSIvA rather than Non-linear ICP on nonlinear MLP graphs | Causal structure induction · MLP synthetic, ER-2, N=10 | Hamming distance: 5.12 vs 9.12 | Improved |
| use CSIvA rather than Non-linear ICP on continuous graphs | Causal structure induction · Continuous synthetic, ER-2, N=10 | Hamming distance: 1.73 vs 2.41 | Improved |
| use CSIvA rather than DAG-GNN on Dirichlet graphs | Causal structure induction · Dirichlet synthetic, ER-2, N=10, α=1.0 | Hamming distance: 4.59 vs 9.3 | Improved |
| train on Dirichlet data with α=5.0 rather than α≤0.5 (ablation) | CSIvA · Dirichlet synthetic, N=10 | Hamming distance: 4.91 vs 0.72 | Hurt |
| use 1500 rather than 1000 samples per CBN (ablation) | CSIvA · Dirichlet synthetic, ER-2 graphs | Hamming distance | Improved |
| use 1000 samples per CBN rather than fewer samples (ablation) | CSIvA · Dirichlet synthetic, ER-1 graphs | Hamming distance | Improved |
| train on ER-2 rather than ER-1 graphs (ablation) | CSIvA on MLP data · MLP synthetic, N=7 | Hamming distance: ER-1: 1.2 -> 3.3; ER-2: 1.8 -> 1.8; ER-3: 5.0 -> 2.8 | Mixed |
| train with Dirichlet α=0.25 rather than α=0.1 (ablation) | CSIvA · Dirichlet synthetic, ER-1, N=7 | Hamming distance: 0.31 -> 0.33; 0.72 -> 0.40; 1.8 -> 0.71 | Mixed |
| increase Dirichlet α from 0.1–0.5 to 1.0 or 5.0 (ablation) | CSIvA · Dirichlet synthetic | Hamming distance: 0.72 -> 2.31; 1.27 -> 4.91 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DAG-GNN — DAG-GNN: DAG Structure Learning with Graph Neural Networks (ICML 2019) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- A Fixed-Point Approach for Causal Generative Modeling (ICML 2024 · similarity 0.90)
- Rethinking Causal Relationships Learning in Graph Neural Networks (AAAI 2024 · similarity 0.88)
- Characterization and Learning of Causal Graphs from Hard Interventions (NeurIPS 2025 · similarity 0.88)
- Estimating Interventional Distributions with Uncertain Causal Graphs through Meta-Learning (NeurIPS 2025 · similarity 0.88)
- Sample Efficient Bayesian Learning of Causal Graphs from Interventions (NeurIPS 2024 · similarity 0.88)
- How Transformers Learn Causal Structure with Gradient Descent (ICML 2024 · 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.
