Learning Structural Causal Models through Deep Generative Models: Methods, Guarantees, and Challenges
Audrey Poinsot, Alessandro Leite, Nicolas Chesneau, Michèle Sébag, Marc Schoenauer
TL;DR
This work addresses the challenge of learning deep structural causal models (DSCMs) from observational data given a known causal graph to answer counterfactual questions. It introduces a two-dimensional taxonomy: (i) classes of deep generative models (invertible explicit, amortised explicit, amortised implicit) and (ii) classes of structural causal models (Bijective Generation Mechanisms, Neural Causal Models), linking each to identifiability guarantees for $\mathcal{L}_2$/$\mathcal{L}_3$ queries. The paper provides a theoretical and empirical synthesis of these methods, highlighting identifiability results, abduction strategies, data assumptions, and handling of hidden confounding, while noting the lack of standardized benchmarks and the need for uncertainty quantification. Practically, it guides practitioners in selecting methods for counterfactual inference in high-stakes domains and outlines open questions around partial identifiability, evaluation benchmarks, and robust uncertainty estimation.
Abstract
This paper provides a comprehensive review of deep structural causal models (DSCMs), particularly focusing on their ability to answer counterfactual queries using observational data within known causal structures. It delves into the characteristics of DSCMs by analyzing the hypotheses, guarantees, and applications inherent to the underlying deep learning components and structural causal models, fostering a finer understanding of their capabilities and limitations in addressing different counterfactual queries. Furthermore, it highlights the challenges and open questions in the field of deep structural causal modeling. It sets the stages for researchers to identify future work directions and for practitioners to get an overview in order to find out the most appropriate methods for their needs.