DAG-WGAN: Causal Structure Learning With Wasserstein Generative Adversarial Networks

TL;DR

DAG-WGAN introduces a novel Wasserstein-based generative adversarial network framework for causal structure learning, effectively capturing data distributions and learning DAGs with improved scalability and handling of diverse data types.

Contribution

The paper presents DAG-WGAN, the first to incorporate Wasserstein distance into causal structure learning with GANs, enhancing data generation and structure discovery.

Findings

Outperforms state-of-the-art models in causal discovery tasks.

Scales well to large datasets and handles both continuous and discrete data.

Achieves accurate causal structure learning with improved data generation.

Abstract

The combinatorial search space presents a significant challenge to learning causality from data. Recently, the problem has been formulated into a continuous optimization framework with an acyclicity constraint, allowing for the exploration of deep generative models to better capture data sample distributions and support the discovery of Directed Acyclic Graphs (DAGs) that faithfully represent the underlying data distribution. However, so far no study has investigated the use of Wasserstein distance for causal structure learning via generative models. This paper proposes a new model named DAG-WGAN, which combines the Wasserstein-based adversarial loss, an auto-encoder architecture together with an acyclicity constraint. DAG-WGAN simultaneously learns causal structures and improves its data generation capability by leveraging the strength from the Wasserstein distance metric. Compared with other models, it scales well and handles both continuous and discrete data. Our experiments have evaluated DAG-WGAN against the state-of-the-art and demonstrated its good performance.