Towards Understanding When Causal Structure Improves Robustness: Evidence from Generative Models

Manal Benhamza, Marianne Clausel, Myriam Tami

Keywords: Causal Generative Models, Variational Autoencoders, Robustness to Distribution Shifts, Structural Causal Models

Abstract: Causal Generative Models aim to incorporate causal knowledge into black box deep generative architectures, thereby improving transparency and interpretability. It is commonly claimed that integrating causal structure also enhances robustness. However, to the best of our knowledge, no prior work has systematically compared the robustness of standard generative models with that of their causal counterparts. Hence, in this work, we aim to address this gap by providing a principled comparative study of robustness between standard and causal generative frameworks, that is theoretically grounded in the VAE setting. Our extensive experiments conducted on synthetic and real-world datasets across different configurations and scales, show that the difference in robustness levels between the standard and causal frameworks is tightly related to the structure of the encoded causal mechanisms. In particular, we provide an intuitive explanation based on the critical properties of the underlying causal graph.

Submission Number: 25