DAGAF: A directed acyclic generative adversarial framework for joint structure learning and tabular data synthesis.

Area of Science:

• Machine Learning
• Causal Inference
• Data Science

Background:

• Causal discovery from observational data is vital for understanding tabular datasets.
• Existing methods often rely on single causal models like Additive Noise Model (ANM) or Linear non-Gaussian Acyclic Model (LiNGAM).
• These single-model approaches may limit the accurate representation of complex data-generating processes.

Purpose of the Study:

• To develop a novel dual-step framework for both causal structure learning and tabular data synthesis.
• To enable learning under multiple causal model assumptions, enhancing flexibility and accuracy.
• To improve the discovery of causal relationships and the generation of realistic synthetic data.

Main Methods:

• Utilizes Directed Acyclic Graphs (DAGs) to model causal relationships among variables.
• Employs a range of functional causal models, including ANM, LiNGAM, and Post-Nonlinear (PNL) models.
• Implicitly learns DAG structures by simulating data generation processes to match real data distributions.

Main Results:

• Achieves significantly lower Structural Hamming Distance (SHD) scores compared to state-of-the-art methods.
• Demonstrates substantial performance improvements on real-world and benchmark datasets (e.g., Sachs: 47%, Child: 11%).
• Successfully generates diverse, high-quality synthetic tabular data samples.

Conclusions:

• The proposed framework effectively integrates causal structure learning and data synthesis.
• It offers superior performance in causal discovery and data generation by leveraging multiple causal models.
• This approach provides a more robust method for understanding and replicating complex data distributions.