Generative Counterfactual Manifold Perturbation: A Robust Framework for Treatment Effect Estimation with Unobserved Confounders

ICLR 2026 Conference Submission13851 Authors

Keywords: ML: Causal Learning

TL;DR: Our method proposes a framework for multi‑treatment effect estimation that can mitigate unobserved confounder and boost accuracy.

Abstract: Estimating treatment effects from observational data is difficult when unobserved confounders create spurious associations that bias simple estimators. Recent generative approaches learn outcome distributions with conditional diffusion models, and some robust representation methods introduce sensitivity analysis or structural priors. These advances work well when identification assumptions hold exactly, but they become fragile when those assumptions are only approximate and offer few practical diagnostics. We introduce Generative Counterfactual Manifold Perturbation (GCMP), a unified framework that combines causal-aware self supervised learning, conditional diffusion counterfactual proxy generation, and adaptive variational inference. GCMP makes three main contributions: (i) a self supervised objective that preserves confounding signals during representation learning; (ii) a conditional diffusion model that reframes proxy construction as a generative task over rich perturbation manifolds; and (iii) an adaptive regularization scheme that yields graceful degradation and calibrated uncertainty when identification assumptions are violated. We also present new identifiability conditions, finite sample error bounds, and diagnostic tests to quantify manifold quality and effective orthogonality. Extensive experiments on synthetic and semi-synthetic benchmarks show that GCMP consistently outperforms the state-of-the-art.

Primary Area: causal reasoning

Submission Number: 13851