Out-of-Sample Tuning for Causal Discovery.

Area of Science:

• Causal inference and machine learning
• Probabilistic graphical models

Background:

• Causal discovery algorithms require hyperparameter tuning, posing a challenge due to the unsupervised nature and unknown true graph.
• Numerous hyperparameter combinations exist for causal graphical probabilistic models.

Purpose of the Study:

• To propose and evaluate Out-of-Sample Causal Tuning (OCT) for selecting optimal hyperparameter combinations in causal discovery.
• To provide a robust tuning method applicable to general causal settings.

Main Methods:

• OCT treats causal models as predictive models, utilizing out-of-sample protocols.
• Employs an information-theoretic approach for mixed data types and a penalty for graph complexity.
• Evaluated using a novel causal-based simulation method for realistic datasets.

Main Results:

• OCT demonstrates strong performance across various experimental scenarios.
• Outperforms other tuning approaches like stability-based and in-sample fitting methods.
• Effectively handles latent confounders and nonlinear relationships.

Conclusions:

• Out-of-Sample Causal Tuning (OCT) is an effective and robust method for hyperparameter optimization in causal discovery.
• The proposed method offers a practical solution for practitioners dealing with complex causal models.