AI Paper Atlas

Research map / Graph learning / Graph algorithms and clustering

Structurally Separated DAG Learning with Multi-Scale Normalized Closure

Shuai Feng, Hao Yan

NeurIPS 2026 · Graph learning › Graph algorithms and clustering · approach: Neural combinatorial optimization on graphs · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
use bounded WG + SMNC closure instead of DAGMADAGMA · linear Gaussian SEM d=100, ER6directed SHD: 88.2 -> 20.8Improved
use bounded WG + SMNC closure instead of DAGMADAGMA · linear Gaussian SEM d=100, ER8directed SHD: 376.8 -> 248.4Improved
use bounded WG + SMNC closure instead of DAGMADAGMA · linear Gaussian SEM d=100, ER2directed SHD: 2.00 vs 2.20No clear effect
use induced + SMNC instead of DAGMADAGMA · nonlinear Gaussian SEM d=100, ER4 and ER8directed SHD: 116.60 -> 85.40 (ER4); 390.40 -> 247.00 (ER8)Improved
use induced + SMNC instead of DAGMADAGMA · nonlinear Gaussian SEM d=100, ER2directed SHD: 68.80 vs 69.80No clear effect
enable SMNC cycle term (ablation)nonlinear backbone · Sachsdirected SHD: 24 -> 14Improved
add hard reverse-edge exclusion from reachability cues (ablation)SMNC · SachsF1-dir: 0.533 -> 0.500Hurt
add positive-only soft reachability cues (ablation)SMNC · Sachsdirected SHD: 14 -> 10Improved
add hard exclusion plus two-sided soft cues (ablation)SMNC · Sachsdirected SHD: 14 -> 6Improved
normalize closure logits by path-length budget (SMNC vs raw head) (ablation)closure head · single cycle L=16, d=32closing-edge gate: 0.1045 -> 0.007290Improved
normalize closure logits by path-length budget (SMNC vs raw head) (ablation)closure head · single cycle L=4 and 8, d=32closing-edge gate: 0.03918 (raw) vs 0.1605 (SMNC) at L=4Hurt
normalize closure logits (SMNC vs raw head) (ablation)closure head · two cycles Ls=8, L_long up to 48cycles broken: raw breaks only L_long=24; SMNC breaks all fourImproved

Experimental setup

Builds on: NOTEARS

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy