Research map / Probabilistic and scientific modeling / Causal discovery
Modular Learning of Deep Causal Generative Models for High-dimensional Causal Inference
Read the paperCheck your idea against 108,967 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 Modular-DCM, an adversarial (WGAN-GP) modular training algorithm for deep causal generative models with latent confounders that trains c-components in an H-graph-derived order, allowing pre-trained generators to be plugged in for identifiable interventional sampling on high-dimensional data.
Problem
Causal inference with high-dimensional variables such as images under latent confounding is impractical because existing identification algorithms need accurate joint distribution estimates, and training deep causal generative models jointly is costly and hard to converge.
Approach
Partition the causal graph into c-components and train them modularly in an order given by an H-graph, using do-calculus rule-2 (modularity condition) to replace interventional samples with observational conditional samples, so that pre-trained conditional generators can be frozen and reused while provably sampling identifiable interventional queries.
Key results
- On the MNIST frontdoor graph, Modular-DCM reaches TVD 0.13 for P(A|do(D=0)) at epoch 300 versus 0.36 for DCM-Rep and 0.43 for NCM.
- Image FID for do(D=0) is 27.20 for Modular-DCM with a pre-trained generator, compared with 80.65 for DCM-Rep and 59.59 for NCM.
- On CelebA-HQ, the Augmented classifier raises accuracy on the Sex=0, Eyeglass=1 subgroup from 0.810 to 0.901.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use modular training instead of training all mechanisms jointly (ablation) | Colored-MNIST frontdoor causal model · Colored-MNIST frontdoor, P(A|do(D=0)) | TVD: 0.36 -> 0.13 | Helped |
| use modular training instead of training all mechanisms jointly (ablation) | Colored-MNIST frontdoor causal model · Colored-MNIST frontdoor generated images | FID: 80.65 -> 27.20 | Helped |
| use a pre-trained model while training all mechanisms together (ablation) | NCM · Colored-MNIST frontdoor generated images | FID: 26.60 -> 192.27 | Hurt |
| use modular training instead of training all mechanisms jointly | Colored-MNIST diamond causal model · Colored-MNIST diamond, I1 | FID: 80.54 -> 32.88 | Helped |
| use modular training instead of training all mechanisms jointly | Colored-MNIST diamond causal model · Colored-MNIST diamond, I2 | FID: 65.17 -> 11.80 | Helped |
| combine the training dataset with generated interventional data | CelebA-HQ eyeglass classifier · CelebA-HQ test domain, Sex=0 | accuracy: 0.948 -> 0.97 | Helped |
| combine the training dataset with generated interventional data | CelebA-HQ eyeglass classifier · CelebA-HQ test domain, Eyeglass=1 | accuracy: 0.847 -> 0.905 | Helped |
| combine the training dataset with generated interventional data | CelebA-HQ eyeglass classifier · CelebA-HQ test domain, Sex=0 and Eyeglass=1 | accuracy: 0.810 -> 0.901 | Helped |
Surprising findings
- Equipping NCM with a pre-trained generator and training all mechanisms jointly worsens FID from 26.60 to 192.27, showing joint training degrades image quality.
Limitations
- Assumes the true semi-Markovian ADMG is fully known.
- Does not handle conditional sampling beyond rejection sampling, which is practical only for low-dimensional evidence.
- Restricted to semi-Markovian models where each confounder affects only two observed variables.
Experimental setup
Datasets: Colored-MNIST CelebA-HQ COVIDx CXR-3
Benchmarks: Colored-MNIST CelebA-HQ COVIDx CXR-3
Metrics: total variation distance KL-divergence FID classifier accuracy
Baselines: NCM DCM-Rep NCM Pre-trained IRM TrainDomain
Builds on: Neural causal models (NCM) Tian's factorization WGAN-GP InterFaceGAN StyleGAN
Closest papers by meaning
- Conditional Generative Models are Sufficient to Sample from Any Causal Effect Estimand (NeurIPS 2024 · similarity 0.90)
- A Fixed-Point Approach for Causal Generative Modeling (ICML 2024 · similarity 0.88)
- Toward Causal Generative Modeling: From Representation to Generation (AAAI 2025 · similarity 0.87)
- Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · similarity 0.86)
- Learning Time-Aware Causal Representation for Model Generalization in Evolving Domains (ICML 2025 · similarity 0.86)
- Neural Causal Abstractions (AAAI 2024 · similarity 0.86)
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.