AI Paper Atlas

Research map / Probabilistic and scientific modeling / Causal discovery

Order-based structure learning for zero-inflated count data under data heterogeneity

Fariha Taskin, Junsouk Choi, Hyunwoong Chang

NeurIPS 2026 · Probabilistic and scientific modeling › Causal discovery · approach: Differentiable and score-based DAG learning

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 an order-based structure learning method for DAGs over zero-inflated multivariate count data, replacing local DAG-space search with an ordering-based scoring framework.

Problem

Causal structure learning for DAG models of zero-inflated multivariate count data; existing zero-inflated Poisson structure learners rely on local search over DAG space, which scales poorly and converges to suboptimal solutions.

Approach

Shifts the search from local DAG-space moves to an order-based scoring framework for zero-inflated Poisson models, exploiting identifiability of ZIP causal models and handling data heterogeneity.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
jointly estimate K=5 datasets instead of K=1 under a shared ordering (ablation)ZIO · Simulation p=40Kendall tau: 0.311 -> 0.740Improved
increase K from 1 to 20 (ablation)ZIO · Simulation p=40edge difference: 7.967 -> 4.062Improved
joint estimation with K=5 at 75% zeros, n=250 (ablation)ZIO · Simulation p=20TPR: 0.597 -> 0.662Improved
use order-based search instead of DAG-space searchZIO vs ZIPBN · Simulation p=75FDR: 0.281 -> 0.139Improved
order-based method vs ZICO at small pZIO vs ZICO · Simulation p=25TPR: 0.938 vs 0.917Hurt
increase zero proportion from 25% to 75% (ablation)ZIO · Simulation p=20, n=250, K=1TPR: 0.903 -> 0.597Hurt
increase sample size from 250 to 2000 at 75% zeros (ablation)ZIO · Simulation p=20, K=1TPR: 0.597 -> 0.937Improved
replace exact with forward-backward parent selection (ablation)ZIO · Exhaustive p=7optimal recovery: 86.0 -> 84.0 (n=500); 94.0 / 94.0 (n=1000)No clear effect
use five random starts instead of one (ablation)R2R hill climbing · Exhaustive p=7, n=250finite-search success: 76.3 -> 95.5Improved
use adjacent transposition instead of R2R (ablation)hill climbing · Simulation p=10, K=1penalized score convergenceHurt
use random transposition instead of R2R (ablation)hill climbing · Simulation p=10, K=1penalized score convergenceNo clear effect
violate common ordering (U<=0.7) in joint estimation (ablation)ZIO K=5 · Simulation p=10, n=1000Hamming distanceHurt

Experimental setup

Builds on: zero-inflated Poisson causal models

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