Evaluation of Causal Structure Learning Methods on Mixed Data Types

摘要

Causal structure learning algorithms are very important in many fields, including biomedical sciences, because they can uncover the underlying causal network structure from observational data. Several such algorithms have been developed over the years, but they usually operate on datasets of a single data type: continuous or discrete variables only. More recently, we and others have proposed new causal structure learning algorithms for mixed data types. However, to-date there is no study that critically evaluates these methods’ performance. In this paper, we provide the first extensive empirical evaluation of several popular causal structure learning methods on mixed data types and in a variety of parameter settings and sample sizes. Our results serve as a guide as to which method performs the best in a given context, and as such they are a first step towards a “method selection guide” for those running causal modeling methods on real life datasets.