Causal discovery for data with continuous and discrete variables in the presence of hidden common causes
Abstract
Machine learning and causal inference have been extensively studied these days, and their cross-fertilization has also been investigated to further advance researches in both of the fields. For example, machine learning techniques are used to better estimate nonlinear causal effects in causal inference, whereas causal knowledge is used for better prediction performance and interpretability of machine learning. In either of the cases, causal structures of variables first need to be specified. In reality, it is common that both of continuous and discrete variables exist in observational data on which one wants to apply causal inference methods and machine learning methods. Moreover, hiddencommon causes are likely to exist in such observational data. Nevertheless, existing causal discovery algorithms and their implementations are able to handle only single type variables, i.e., either of continuous variables or discrete variables, and do not allow mixtures of continuous and discrete variables.Thus, we aim to develop causal discovery methods to estimate causal structures from data in which continuous and discrete variables coexist and hidden common causes exist. An important point of this project is that we do this in the presence of hidden common causes. Removing or alleviating the effect of hidden common causes is one of the biggest challenges when making causal inference from observational data. Our approach would give the very first method capable of uniquely estimating causal structures of mixtures of continuous and discrete variables in the presence of hidden common causes. We also demonstrate the effectiveness of such causal discovery algorithms in causal effect identification problems and prediction problems.We would publish journal/conference papers and make conference presentations to present research results. We are also planning to implement our methods and distribute the codes on the web.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 20, 2020
- Source ID
- N000142012501
Entities
People
- Shohei Shimizu
Organizations
- Office of Naval Research
- Shiga University
- United States Navy