Abstract
Estimating heterogeneous causal effects under complex exposure patterns is an important yet challenging problem in modern data analysis. Conventional double machine learning methods provide a principled framework for debiased causal estimation, but they typically rely on manually specified exposure representations and may perform inadequately when exposure information is high-dimensional, continuous, or structurally complex. To address this limitation, we propose EEDML-HTE, a fuzzy Bayesian exposure modeling framework for heterogeneous causal effect estimation. The key idea is to learn a low-dimensional exposure embedding that captures latent similarity and structural variation in observed exposure patterns, and then characterize local exposure neighborhoods through fuzzy membership weights within an orthogonal-score-based double machine learning procedure. The learned embedding is intended to preserve treatment-effect-relevant variation while compressing redundant exposure details, so that nearby points in the latent space remain causally comparable rather than merely numerically similar. The proposed modeling strategy integrates representation learning with cross-fitting, Neyman-orthogonal score construction, and Bayesian regularization on the second-stage heterogeneous effect learner, allowing flexible nuisance estimation while mitigating regularization bias. By leveraging the learned exposure embedding, fuzzy exposure assignment, and posterior-stabilized effect learning, EEDML-HTE can improve the identification of treatment effect heterogeneity across individuals with comparable latent exposure profiles. We further develop the corresponding theoretical modeling framework for conditional average treatment effects and discuss the conditions under which representation compression remains valid, while empirical results illustrate finite-sample behavior across multiple public data-generating settings. These results suggest that combining fuzzy exposure modeling, Bayesian regularization, and double machine learning provides a promising direction for heterogeneous causal inference in high-dimensional and structurally rich environments.