CausalStructCodec: Causally-aware observational and interventional data generator

Louis Hernandez, Matthieu Boussard

Keywords: Generative Models, Codecs, Causality, Interventions

TL;DR: We present a novel causally-aware architecture to generate interventional and observational

Abstract: Over the last few years, causal generative models have massively gained popularity. Their main goal is to generate observational, interventional and counterfactual data. They are also interesting for causal discovery or fair Machine Learning. These generators are based on typical data generation architecture, such as VAEs, GANs or normalizing flows. However, a recent data generation architecture, the Codecs, is more computationally efficient and allows for complex data generation. In this work, we introduce the CausalStructCodecs (CSC), a novel causally-aware architecture based on a specific kind of codec, the StructCodec. We show that results for non-complex data are level with state-of-the-art models for observational and interventional data generation, in significantly fewer epochs.