Generative Causal Structure Learning with Dual Latent Spaces and Annealing

Abstract

In this work, we address causal structure learning in the presence of unobserved confounders. Such causal structures can be represented by Acyclic Directed Mixed Graphs (ADMGs), where observed cause-effect relations are depicted by directed edges and unobserved confounded relations by bidirected edges. Prior methods for causal structure learning with unobserved common causes have primarily focused on search-based approaches, and more recently on flow-based generative models. We propose a novel generative method based on a variant of the Variational Autoencoder (VAE) with dual latent spaces to represent the directed cause-effect relations and the bidirected unobserved confounded relations, associating two trainable adjacency matrices. To enhance the learning process, we introduce a causality constraint combined with the concept of a causal annealing strategy during training, guiding the learning toward meaningful causal structures. Experimental results show that our method achieves competitive performance in identifying both observed and latent causal relationships on synthetic datasets. Furthermore, we demonstrate that the learned causal structure significantly improves downstream causal inference performance on real-world data.