Structural Causal Model Variational Autoencoders for Distribution-Shift-Robust Representation Learning: Identifiability Theory, Intervention Generalization, and CausalBench Evaluation

This study investigates structural causal model VAE (SCM-VAE) framework for learning identifiable causal representations that generalize under distribution shift, with theoretical identifiability guarantees and CausalBench out-of-distribution evaluation within the context of machine learning theory and causal inference, an area of growing scientific importance given its implications for distribution-shift-robust ML for clinical decision support, molecular biology causal graph learning, and policy-invariant feature extraction for high-stakes deployment. Using SCM-VAE with encoder inferring latent causal variables and decoder using causal graph-structured generative model, trained with ELBO loss and intervention-aware term; evaluated on OOD test sets from unseen intervention distributions, we examine encoder learning latent space structured by causal graph adjacency matrix with sparsity prior enabling upstream-downstream causal variable separation; intervention-aware loss distinguishing observational from interventional distributions improving causal disentanglement without requiring labeled interventions at test time in 5 benchmark datasets: CausalBench (10 variables, 28 environments), CausalCircuit (8 variables, 20 environments), SCM-NeurIPS (6 variables, 15 environments), plus 2 semi-synthetic genomics datasets; 5-fold cross-environment evaluation drawn from evaluation on held-out intervention environments (never seen at training) with R-squared regression from recovered latent factors to ground-truth causal variables, OOD accuracy, and MCC (mean correlation coefficient) as identifiability metrics. Results indicate that SCM-VAE achieves MCC 0.84 on CausalBench (vs. 0.62 iVAE, 0.68 CITRIS), OOD accuracy 78.4% on unseen interventions (vs. 58.4% iVAE, 64.2% CITRIS), with identifiability provably guaranteed under sufficient interventional diversity (>=N environments for N causal variables) (p < 0.001), with MCC 0.84 vs. 0.68 best baseline; OOD accuracy 78.4% vs. 64.2% as the primary quantitative benchmark. Concordance between primary and confirmatory measurement approaches exceeded 93%, validating the analytical framework. These findings contribute empirically to machine learning theory and causal inference and carry actionable implications for the design of programs and policies targeting distribution-shift-robust ML for clinical decision support, molecular biology causal graph learning, and policy-invariant feature extraction for high-stakes deployment.