Research map / Probabilistic and scientific modeling / Causal discovery
Causal Representation Learning via Counterfactual Intervention
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 CFI-VAE, a VAE-based causal disentanglement framework that uses counterfactual intervention and a reweighted loss to remove inductive and dataset biases when learning causal effects among annotated factors.
Problem
Existing causal representation learning methods learn biased causal effects among semantic factors because inductive biases (confounding from the image and unobserved factors) and dataset biases (class imbalance) are ignored.
Approach
Estimate causal effects as total direct effect via counterfactual intervention (factual minus counterfactual logits) so inductive-bias effects cancel, and apply class-balanced reweighting to the intervention loss to remove dataset bias; the resulting causal matrix is injected into a nonlinear SCM within a VAE.
Key results
- CFI-VAE reaches MIC 0.872 and TIC 10.187 on average across Age, Smile, Pend and Flow, outperforming CausalVAE-D, CausalGAN and DEAR.
- CFI-VAE achieves 100% ISR and CERD with 0% SCR on all four datasets in the main comparison.
- With 10% supervision, CFI-VAE averages MIC 0.749 and TIC 7.944, still ahead of baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove inductive bias during causal-effect learning using counterfactual intervention (ablation) | CFI-VAE · Age, Smile, Pend, Flow | MIC: 0.615 -> 0.626; 0.851 -> 0.873; 0.785 -> 0.997; 0.973 -> 0.992 | Helped |
| remove dataset bias by reweighting the causal loss (ablation) | CFI-VAE · Age, Smile | CERD: 25 -> 100; 25 -> 100 | Helped |
| remove dataset bias by reweighting the causal loss (ablation) | CFI-VAE · Age, Smile | MIC: 0.619 -> 0.626; 0.860 -> 0.873 | No clear effect |
| use CFI-VAE instead of the strongest baseline | causal representation learning · Age, Smile, Pend, Flow | MIC: 0.609 -> 0.626; 0.802 -> 0.873; 0.951 -> 0.997; 0.773 -> 0.992 | Helped |
| use CFI-VAE instead of the strongest baseline | causal representation learning · Age, Smile, Pend, Flow | TIC: 6.427 -> 7.495; 8.973 -> 10.429; 10.845 -> 11.534; 8.192 -> 11.291 | Helped |
Surprising findings
- Removing only dataset bias keeps MIC/TIC/ISR close to the full model but weakens CERD, suggesting the two biases affect different aspects of causality.
Limitations
- Relies on annotated concepts and a given causal graph, so it requires full supervision.
- Performance drops substantially when supervision proportion is reduced to 10% or lower.
- Evaluated only on CelebA and two small synthetic datasets.
Experimental setup
Datasets: CelebA Pendulum (synthetic) Flow (synthetic)
Benchmarks: CelebA Age CelebA Smile Pendulum Flow
Metrics: MIC TIC ISR CERD SCR
Baselines: CausalVAE CausalVAE-D CausalGAN DEAR
Builds on: CausalVAE DEAR Total Direct Effect class-balanced loss
Closest papers by meaning
- Disentangled Representation Learning in Non-Markovian Causal Systems (NeurIPS 2024 · similarity 0.91)
- Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · similarity 0.90)
- A Sparsity Principle for Partially Observable Causal Representation Learning (ICML 2024 · similarity 0.90)
- Causal Representation Learning from Multiple Distributions: A General Setting (ICML 2024 · similarity 0.89)
- Causal Representation Learning Made Identifiable by Grouping of Observational Variables (ICML 2024 · similarity 0.89)
- Identifiability Guarantees for Causal Disentanglement from Purely Observational Data (NeurIPS 2024 · similarity 0.89)
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.