Research map / Probabilistic and scientific modeling / Causal discovery
Differentiable Constraint-Based Causal Discovery
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 differentiable d-separation scores via percolation theory and soft logic (LogLTN), instantiated as DAGPA, a gradient-based method that optimizes low-order conditional independence constraints for causal discovery in low-sample regimes.
Problem
Causal discovery from observational data is hard with small sample sizes: constraint-based methods rely on unreliable CI tests, while score-based methods typically forgo explicit CI testing.
Approach
Relaxes d-separation/d-connection into differentiable functions over a weighted adjacency matrix using percolation-based reachability and LogLTN soft logic operators, yielding lower bounds on expected d-separation statements, and optimizes them against p-values from CI tests (with acyclicity via DAGMA's log-det regularizer).
Key results
- On Sachs, DAGPA achieves CI-MCC of 0.75-0.98 while all baselines cluster between 0.3 and 0.65.
- In low-sample (n=100) synthetic binary settings, DAGPA outperforms most baselines and is on par with DAGMA-nonlinear, while beating PC and k-PC.
- At n=10000, DAGPA stays competitive with constraint-based methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DAGPA instead of baseline causal discovery methods | Sachs · Sachs | CI-MCC: 0.3-0.65 -> 0.75-0.98 | Helped |
| use DAGPA instead of most baselines | causal discovery methods · Synthetic binary data, n=100 | CI-MCC | Helped |
| use DAGPA instead of DAGMA-Nonlinear | causal discovery methods · Synthetic binary data, n=100 | CI-MCC | No clear effect |
| use DAGPA instead of constraint-based methods | causal discovery methods · Synthetic binary data, n=10000 | CI-MCC | No clear effect |
Limitations
- Relies on p-values with a constant-minus-p heuristic to measure dependence strength.
- Focuses on only low-order (|Z|<=1) conditioning sets and assumes no latent confounders.
- Scalability not yet addressed.
Experimental setup
Datasets: Sachs synthetic binary data from ER and SF DAGs (pyAgrum)
Benchmarks: Sachs synthetic binary ER and SF datasets
Metrics: CI-MCC (Matthews correlation coefficient over low-order CI s graph F1 SHD
Baselines: PC k-PC GES NOTEARS (linear) NOTEARS (nonlinear) DAGMA (linear) DAGMA (nonlinear)
Builds on: NOTEARS DAGMA PC algorithm k-PC LogLTN PCGrad Discrete Langevin Proposal
Closest papers by meaning
- THE ROBUSTNESS OF DIFFERENTIABLE CAUSAL DISCOVERY IN MISSPECIFIED SCENARIOS (ICLR 2025 · similarity 0.90)
- Stable Differentiable Causal Discovery (ICML 2024 · similarity 0.90)
- On Causal Discovery in the Presence of Deterministic Relations (NeurIPS 2024 · similarity 0.89)
- Causal Discovery with Fewer Conditional Independence Tests (ICML 2024 · similarity 0.89)
- Foundations of Testing for Finite-Sample Causal Discovery (ICML 2024 · similarity 0.89)
- On the sample complexity of conditional independence testing with Von Mises estimator with application to causal discovery (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.