Evaluation and Comparison of Causal Discovery Algorithms

Causal discovery algorithms are very important in many fields because of their ability to reveal the underlying causal structure from observational data. With more and more causal discovery algorithms being developed, it is hard for researchers to determine which algorithm to use for their specific domain of work. In this paper, several causal discovery algorithms are made to run across datasets with varying variable numbers and observation numbers, then their performance is evaluated using the same metric. The result shows a low varsortability, such as 0.21 is achievable with synthetic datasets. Algorithms that rely on the assumption of causal sufficiency may exhibit similar performance on datasets that either with or without hidden variables, provided that the number of such variables is minimal. Algorithms that assume acyclicity, on the other hand, may display instability in their performance when applied to datasets containing cycles. Regarding the overall performance of tested causal discovery algorithms, it has been observed that GES underperformed
in comparison to other algorithms. Therefore, it may not be considered the optimal choice for causal discovery. ... mehrIn this paper, the performance of several causal discovery algorithms is evaluated. Future research will focus on gathering more algorithms and datasets to do more cross-comparison in order to better understand the practical characteristics of di↵erent causal discovery algorithms.