AI Paper Atlas

Research map / Probabilistic and scientific modeling / Causal discovery

SCDM: Scalable Causal Discovery in Nonlinear Temporal Systems with Meta-Learning

Jingbo Wang, Kegeng Tang, Zihao Wang, Shaogang Ren

NeurIPS 2026 · Probabilistic and scientific modeling › Causal discovery · approach: Causal discovery from time series · 1 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 SCDM, a shared-parameter meta-learning framework for scalable causal discovery in high-dimensional nonlinear multivariate time series.

Problem

Causal discovery from nonlinear multivariate time series in high-dimensional systems, where candidate directed relations grow quadratically with the number of variables; existing methods suffer from target-wise fitting, repeated conditional testing, or expensive graph search.

Approach

A shared-parameter meta-learning framework that performs temporal causal discovery in a scalable way, avoiding per-target model fitting, repeated conditional tests, and graph-search procedures.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Subsample target tasks B=32 instead of all targets (B=100) (ablation)SCDM · Simulated fMRI 100 variablestraining time (s): 2049.9 -> 668.4Improved
Subsample B=32 instead of B=100 (ablation)SCDM · Simulated fMRI 100 variablesAUPRC: 0.552 -> 0.728Improved
Reduce task batch from 32 to 10 (ablation)SCDM · Simulated fMRI 100 variablesAUPRC: 0.728 -> 0.656Hurt
Use support/query meta-objective instead of pooled multi-task training (ablation)SCDM · Held-out target adaptationquery MSE: 4.305 -> 3.782Improved
Add inner adaptation steps (K>0) versus K=0 (ablation)SCDM · Held-out target adaptationquery MSE: 5.487 -> 3.782Improved
Add shared causal-strength matrix W (ablation)SCDM · Simulated fMRIAUPRC: 0.390 -> 0.728Improved
Use structural readout instead of source-isolated readout (ablation)SCDM · Simulated fMRIAUPRC: 0.693 -> 0.728Improved
Use source-isolated readout instead of structural readout (ablation)SCDM · Simulated EEGAUPRC: 0.426 -> 0.476Improved
Use SCDM instead of UnCLe on NC8SCDM · NC8AUROC: 0.947 -> 0.916Hurt
Use SCDM instead of GVAR on simulated EEGSCDM · Simulated EEGAUROC: 0.750 -> 0.717Hurt
Use SCDM instead of UnCLe on ROI simulated fMRISCDM · Simulated fMRI (ROI-level)AUROC: 0.661 -> 0.867Improved
Use SCDM on 1000-variable fMRI at 200 s budgetSCDM · Voxel-level fMRI (d=1000), 200 sAUPRC enrichment: 2.05x -> 1.96xHurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most 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

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