Research map / Graph learning / Graph algorithms and clustering
Revisiting Differentiable Structure Learning: Inconsistency of L1 Penalty and Beyond
Kaifeng Jin, Ignavier Ng, Kun Zhang, Biwei Huang
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 |
|---|---|---|---|
| add the estimated moral graph to restrict the search space (ablation) | CALM (hard constraint) · 50-node ER1 (infinite samples, standardized) | SHD of CPDAG: 16.7 -> 5.5 | Improved |
| add the estimated moral graph to restrict the search space (ablation) | soft-constraint ℓ0 likelihood · 50-node ER1 (infinite samples, standardized) | SHD of CPDAG: 33.8 -> 7.6 | Improved |
| use hard instead of soft DAG constraint (ablation) | ℓ0 likelihood with moral graph · 50-node ER1 (infinite samples, standardized) | SHD of CPDAG: 7.6 -> 5.5 | Improved |
| use hard instead of soft DAG constraint (no moral graph) (ablation) | ℓ0 likelihood without moral graph · 50-node ER1 (infinite samples, standardized) | SHD of CPDAG: 33.8 -> 16.7 | Improved |
| standardize the data before CALM (ablation) | CALM · 50-node ER1 (infinite samples) | SHD of CPDAG: 9.9 -> 5.5 | No clear effect |
| replace ℓ1-penalized GOLEM-NV with CALM (ℓ0 Gumbel-Softmax, hard DAG, moral graph) | GOLEM-NV-ℓ1 · 50-node ER1 (n=1000, standardized) | SHD of CPDAG: 60.0 -> 12.1 | Improved |
| replace ℓ1-penalized GOLEM-NV with CALM | GOLEM-NV-ℓ1 · 50-node ER4 (n=1000, standardized) | SHD of CPDAG: 211.6 -> 168.8 | Improved |
| replace ℓ1-penalized GOLEM-NV with CALM | GOLEM-NV-ℓ1 · Sachs real data | SHD of CPDAG: 13 -> 12 | Improved |
| use CALM instead of discrete PC/FGES on sparse graphs | 50-node ER1 (n=1000) · 50-node ER1 (n=1000, standardized) | SHD of CPDAG: CALM 12.1 vs PC 11.0, FGES 8.4 | Hurt |
| use CALM instead of PC/FGES on dense graphs | 50-node ER4 (n=1000) · 50-node ER4 (n=1000, standardized) | SHD of CPDAG: PC 200.5 / FGES 425.6 -> CALM 168.8 | Improved |
| use ℓ1 instead of ℓ0 penalty (minimum-ℓ1 DAG among covariance-consistent DAGs) (ablation) | linear Gaussian structure learning · 8-node ER1 population covariance | number of edges / SHD of CPDAG: edges 8.0 -> 22.74; SHD 0 -> 19.97 | Hurt |
Closest papers by meaning
- Markov Equivalence and Consistency in Differentiable Structure Learning (NeurIPS 2024 · similarity 0.94)
- Differentiable Structure Learning with Partial Orders (NeurIPS 2024 · similarity 0.90)
- Differentiable Structure Learning and Causal Discovery for General Binary Data (NeurIPS 2025 · similarity 0.88)
- Differentiable Structure Learning with Ancestral Constraints (ICML 2025 · similarity 0.88)
- Scores for Learning Discrete Causal Graphs with Unobserved Confounders (AAAI 2024 · similarity 0.87)
- Boosting Causal Structure Learning: An Asymmetric Exponential Modulation Gaussian-Based Adaptive Sample Reweighting Framework (AAAI 2025 · similarity 0.85)
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.
