On causal structural learning algorithms: Oracles' simulations and considerations

This work evaluates the performance of several causal structure learning algorithms, in terms of their effectiveness and efficiency in detecting true causal relations among variables. Constraint-based, score -based and hybrid algorithms are jointly compared and ranked according to the two criteria above and their performance is evaluated when used in either directed or undirected acyclic graphs. Fixing the number of variables considered, a Monte Carlo simulation is run for constructing linear causal effects among variables, both in small and large data samples with different causal network properties. Latent confounding variables are empirically demonstrated to be the main drawback of an algorithms' performance, independently of the size of the sample. & COPY; 2023 Elsevier B.V. All rights reserved.

Automated summary

This study evaluates the performance of various causal structure learning algorithms in detecting true causal relations among variables. Constraint-based, score-based, and hybrid algorithms are compared and ranked based on effectiveness and efficiency in directed or undirected acyclic graphs. Monte Carlo simulations are conducted to create linear causal effects among variables in different sample sizes and causal network properties. The presence of latent confounding variables is found to be the main limitation of the algorithms, regardless of the sample size.