University of Strathclyde
Causal discovery from observational tabular data with generative adversarial learning
Abstract
dc:description.abstractBackgroundCausal knowledge is essential for understanding complex systems and revealing relationships between variables. It enables researchers to transition beyond correlations,reason about cause and effect, and derive scientific insights. Although RandomizedControlled Trials (RCT) remain the gold standard for causal inference, they are ofteninfeasible due to ethical, logistical, or financial constraints and may lack real-worldapplicability. In contrast, observational data offer abundant, diverse samples, making them well-suited for large-scale analysis. Despite susceptibility to confounding,advances in structure learning from observations allow researchers to identify causalrelationships without relying on randomized experiments.Research objectivesThis thesis challenges conventional maximum likelihood estimation (MLE)-based methods by exploring adversarial causal discovery approaches. It leverages the WassersteinGenerative Adversarial Network with Gradient Penalty (WGAN-GP) framework toaddress key limitations: (1) model overfitting from simplistic loss functions; (2) dependence on single parametric assumptions that hinder accurate causal graph recoveryreflective of true data relationships; (3) high computational cost from Augmented Lagrangian optimization in the NOTEARS framework; and (4) inability to perform causaldiscovery and tabular data synthesis simultaneously under a single framework.MethodsThree models were developed using the WGAN-GP framework. The first, DAG-WGANintegrates WGAN-GP with variational inference, leveraging hybrid losses for improvedcausal modeling. The second, DAG-WGAN+ enhances continuous optimization withefficient structure learning techniques. The third, DAGAF captures variable interdependencies under various causal assumptions to generate synthetic data preservingcausal relations.ResultsAll models target multivariate causal discovery and were rigorously evaluated usingStructural Hamming Distance (SHD). Results show they outperform leading methodsin causal discovery across 97.47% of all test cases. In real-world experiments, theproposed models achieve superior accuracy (SHD = 8 vs. > 10 for state-of-the-artmodels). Findings further reveal that precise causal modeling enhances synthetic dataquality by preserving underlying causal mechanisms.
Degree
thesis:*- Name dc:type.qualificationname
- phd
- Level dc:type.qualificationlevel
- doctoral-pg
- Grantor dc:publisher.institution
- University of Strathclyde
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Petkov, Hristo
- Advisors dc:contributor.advisor
-
- Dong, Feng
- Maguire, Roma
- Author Identifier dc:creator.authoridentifier
-
- 202092048
Rights
- Licence dc:rights.uri
Identifiers
dc:identifier.*- Identifier
- T17565