Research map / Probabilistic and scientific modeling / Causal discovery
Counterfactual Generative Modeling with Variational 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 Variational Causal Inference (VCI), a variational Bayesian framework that derives an ELBO for individual-level counterfactual outcome likelihood, enabling end-to-end counterfactual supervision and disentangled exogenous noise abduction for high-dimensional counterfactual generation.
Problem
Predicting high-dimensional individual counterfactual outcomes (e.g., gene expression, facial images) under counterfactual treatments when covariates are limited; prior conditional VAE-based methods optimize marginal likelihoods and lack counterfactual supervision and latent disentanglement.
Approach
Replace the conditional VAE objective with an ELBO on the individual-level counterfactual likelihood p(Y'|Y,X,T,T'), with a KL term aligning the latent posterior of factual and counterfactual outcomes; counterfactual supervision uses an adversarial discriminator fit to covariate-specific outcome distributions.
Key results
- On Marson DE genes, VCI reaches R^2 0.658 ± 0.040 versus 0.549 for CPA.
- On Morpho-MNIST, VCI reduces image MSE (x10^-2) to 0.42 for do(thickness) versus 2.20 for the best MED variant.
- On Sciplex DE genes, VCI achieves 0.496 ± 0.011 vs 0.474 for CPA; CPA is slightly higher on Sciplex all genes (0.836 vs 0.832).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use VCI instead of CPA | single-cell perturbation prediction · Sciplex OOD, DE genes | R²: 0.474 ± 0.014 -> 0.496 ± 0.011 | Helped |
| Use VCI instead of CPA | single-cell perturbation prediction · Sciplex OOD, all genes | R²: 0.836 ± 0.002 -> 0.832 ± 0.008 | Hurt |
| Use VCI instead of CPA | single-cell perturbation prediction · Marson OOD, all genes | R²: 0.876 ± 0.005 -> 0.891 ± 0.007 | Helped |
| Use VCI instead of CPA | single-cell perturbation prediction · Marson OOD, DE genes | R²: 0.549 ± 0.019 -> 0.658 ± 0.040 | Helped |
| Use VCI instead of MED (β=3) | counterfactual image generation · Morpho-MNIST do(th) | Image MSE (· 10⁻²): 2.20 -> 0.42 | Helped |
| Use VCI instead of MED (β=3) | counterfactual image generation · Morpho-MNIST do(in) | Image MSE (· 10⁻²): 0.61 -> 0.13 | Helped |
| Use VCI instead of MED (β=3) | counterfactual image generation · Morpho-MNIST mix | Image MSE (· 10⁻²): 1.39 -> 0.36 | Helped |
| Add latent divergence to SAE (ablation) | VCI on Morpho-MNIST · Morpho-MNIST do(th) | Image MSE (· 10⁻²): 0.67 -> 0.42 | Helped |
| Add latent divergence to SAE (ablation) | VCI on Morpho-MNIST · Morpho-MNIST do(in) | Image MSE (· 10⁻²): 0.20 -> 0.13 | Helped |
| Add latent divergence to SAE (ablation) | VCI on Morpho-MNIST · Morpho-MNIST mix | Image MSE (· 10⁻²): 0.47 -> 0.36 | Helped |
| Add latent divergence to SAE (ablation) | VCI on Morpho-MNIST · Morpho-MNIST do(th) | Thickness MAE: 0.32 -> 0.20 | Helped |
| Add latent divergence to SAE (ablation) | VCI on Morpho-MNIST · Morpho-MNIST do(in) | Intensity MAE (· 10⁻¹): 0.26 -> 0.22 | Helped |
Surprising findings
- Diffusion-based counterfactual models (Diff-SCM) underperform simpler HVAE baselines in counterfactual fidelity, which the authors attribute to the conditional VAE formulation discarding individuality.
Limitations
- Theoretical results assume no unobserved confounders.
- Does not explicitly model causal relations among multiple treatments.
- Counterfactual truth unavailable for real datasets such as CelebA-HQ, so quality is assessed only qualitatively.
Experimental setup
Datasets: Sciplex Marson CRISPRa Morpho-MNIST CelebA-HQ
Benchmarks: Sciplex Marson Morpho-MNIST CelebA-HQ
Metrics: R^2 (all genes) R^2 (DE genes) image MSE thickness MAE intensity MAE
Baselines: AE CEVAE GANITE CPA DEAR Diff-SCM CHVAE MED SAE HAE
Builds on: conditional VAE CEVAE CHVAE GANITE AIPW estimator
Closest papers by meaning
- Counterfactual Image Editing (ICML 2024 · similarity 0.90)
- Towards Characterizing Domain Counterfactuals for Invertible Latent Causal Models (ICLR 2024 · similarity 0.89)
- Benchmarking Counterfactual Image Generation (NeurIPS 2024 · similarity 0.88)
- Conditional Counterfactual Mean Embeddings: Doubly Robust Estimation and Learning Rates (NeurIPS 2026 · similarity 0.88)
- Counterfactual Contrastive Learning with Normalizing Flows for Robust Treatment Effect Estimation (ICML 2025 · similarity 0.88)
- Learning Counterfactual Outcomes Under Rank Preservation (NeurIPS 2025 · similarity 0.88)
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.