Deep Causal Generative Modeling for Tabular Data Imputation and Intervention

Xiaohan Jiang, Sen Liu, Hongbin Zhu

Keywords: tabular data, generative models, missing value imputation, causal knowledge

Abstract: Tabular data synthesis could overcome the tabular data incompleteness and data availability issue. In most prior works, deep generative models are basically constructed following standard architecture designs. However, these works do not consider the inter-relationships among the features, or the latent variables. To fully leverage these inter-relationships, we develop a novel causal-aware asymmetric variational autoencoder architecture (CAT) for tabular data generation, imputation, and intervention. The developed model, called CAT-MIWAE, learns exogenous causal representation with a pre-defined causal graph in incomplete data context. It provides interpretability for partially observed features and could efficiently address missing value imputation problem. Besides, CAT-MIWAE can sample data from distributions under arbitrary conditions and interventions. This merit enables us to actively generate counterfactuals or debiased fair data samples for any subpopulation of interest. To validate the effectiveness of the proposed causally aware models, we conduct extensive experiments on real-world tabular datasets. Experiments show that the proposed models outperform the state of the art models. Moreover, we perform CATE estimations to show that CAT-MIWAE model could appropriately extrapolate any conditional or interventional distributions from the original observed data distribution.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Generative models