Dual-Latent Generative Causal Structure Learning with Causal Annealing

We explore causal structure learning with unobserved confounders, represented by Acyclic Directed Mixed Graphs, where directed edges indicate observed cause-effect relationships and bidirected edges capture unobserved confounding. Previous methods have focused on search-based approaches or flow-based generative models.In contrast, we propose a novel variational autoencoder framework with dual latent spaces, each associated with a trainable adjacency matrix to capture directed and bidirected edges, respectively. We propose a causality constraint and introduce a causal annealing strategy during training to obtain meaningful causal graph structures.Experiments show competitive identification of both relationship types on synthetic data, with learned structures enhancing downstream causal inference in a real-world task.