Research map / Graph learning / Graph algorithms and clustering
Structurally Separated DAG Learning with Multi-Scale Normalized Closure
Shuai Feng, Hao Yan
Read the paperCheck your idea against 137,217 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 a structurally separated continuous DAG learner that decouples direct-edge support from signed effect magnitude, with a multi-scale normalized closure for higher-order structural information.
Problem
Continuous DAG learners optimize a single weighted adjacency matrix, so data fitting, sparsity, acyclicity, and auxiliary higher-order structural information (reachability, ancestral order) all act on the same variables.
Approach
Separates the direct-edge support variables from the signed effect magnitudes instead of one weighted adjacency matrix, and uses a multi-scale normalized closure so higher-order relations such as reachability or ancestral order can be imposed without acting directly on the edge variables.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use bounded WG + SMNC closure instead of DAGMA | DAGMA · linear Gaussian SEM d=100, ER6 | directed SHD: 88.2 -> 20.8 | Improved |
| use bounded WG + SMNC closure instead of DAGMA | DAGMA · linear Gaussian SEM d=100, ER8 | directed SHD: 376.8 -> 248.4 | Improved |
| use bounded WG + SMNC closure instead of DAGMA | DAGMA · linear Gaussian SEM d=100, ER2 | directed SHD: 2.00 vs 2.20 | No clear effect |
| use induced + SMNC instead of DAGMA | DAGMA · nonlinear Gaussian SEM d=100, ER4 and ER8 | directed SHD: 116.60 -> 85.40 (ER4); 390.40 -> 247.00 (ER8) | Improved |
| use induced + SMNC instead of DAGMA | DAGMA · nonlinear Gaussian SEM d=100, ER2 | directed SHD: 68.80 vs 69.80 | No clear effect |
| enable SMNC cycle term (ablation) | nonlinear backbone · Sachs | directed SHD: 24 -> 14 | Improved |
| add hard reverse-edge exclusion from reachability cues (ablation) | SMNC · Sachs | F1-dir: 0.533 -> 0.500 | Hurt |
| add positive-only soft reachability cues (ablation) | SMNC · Sachs | directed SHD: 14 -> 10 | Improved |
| add hard exclusion plus two-sided soft cues (ablation) | SMNC · Sachs | directed SHD: 14 -> 6 | Improved |
| normalize closure logits by path-length budget (SMNC vs raw head) (ablation) | closure head · single cycle L=16, d=32 | closing-edge gate: 0.1045 -> 0.007290 | Improved |
| normalize closure logits by path-length budget (SMNC vs raw head) (ablation) | closure head · single cycle L=4 and 8, d=32 | closing-edge gate: 0.03918 (raw) vs 0.1605 (SMNC) at L=4 | Hurt |
| normalize closure logits (SMNC vs raw head) (ablation) | closure head · two cycles Ls=8, L_long up to 48 | cycles broken: raw breaks only L_long=24; SMNC breaks all four | Improved |
Experimental setup
Builds on: NOTEARS
Closest papers by meaning
- DAGs with No Curl: An Efficient DAG Structure Learning Approach (ICML 2021 · similarity 0.89)
- On the Role of Sparsity and DAG Constraints for Learning Linear DAGs (NeurIPS 2020 · similarity 0.88)
- DAG-GNN: DAG Structure Learning with Graph Neural Networks (ICML 2019 · similarity 0.88)
- CoLiDE: Concomitant Linear DAG Estimation (ICLR 2024 · similarity 0.87)
- Analytic DAG Constraints for Differentiable DAG Learning (ICLR 2025 · similarity 0.86)
- 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.
