Reliable ABC model choice via random forests.

Area of Science:

• Computational Statistics
• Machine Learning
• Population Genetics

Background:

• Approximate Bayesian computation (ABC) is used for Bayesian inference in complex models.
• Standard ABC techniques may poorly evaluate model posterior probabilities.
• Model selection in complex scenarios remains a challenge.

Purpose of the Study:

• To develop a novel approach for Bayesian model selection using random forests (RF) within the ABC framework.
• To improve the accuracy, robustness, and efficiency of model selection for complex models.
• To reframe Bayesian model selection as a classification problem.

Main Methods:

• Integration of random forests (RF) into Approximate Bayesian computation (ABC) for model selection.
• Utilizing RF for initial model prediction and subsequent posterior probability approximation.
• Application to controlled experiments and population genetics datasets.

Main Results:

• The ABC RF approach demonstrates improved discriminative power among competing models.
• Enhanced robustness against the choice of data-summarizing statistics.
• Significant reduction in computational effort, achieving at least 50% efficiency gain.
• Accurate approximation of the posterior probability for the selected model.

Conclusions:

• The ABC RF methodology extends the applicability of ABC to larger datasets and more complex models.
• The R package 'abcrf' is available for implementing the proposed methodology.
• This approach offers a powerful new tool for Bayesian model selection in various scientific fields.