Bayesian Additive Regression Trees using Bayesian Model Averaging.

Area of Science:

• Computational Statistics
• Bioinformatics
• Machine Learning

Background:

• Bayesian Additive Regression Trees (BART) is a powerful statistical model but can be computationally expensive for high-dimensional datasets (large number of variables, *p*).
• Random forests are popular for high-dimensional data but lack probabilistic predictions.
• There is a need for efficient algorithms that can handle high-dimensional data and provide probabilistic estimates.

Purpose of the Study:

• To propose BART-BMA, a novel fitting algorithm for BART that enhances efficiency in high-dimensional settings.
• To leverage Bayesian Model Averaging and greedy search for faster posterior distribution estimation.
• To offer a robust model-based approach for analyzing small *n* large *p* datasets.

Main Methods:

• Developed BART-BMA, integrating Bayesian Model Averaging with a greedy search strategy.
• Employed a combination of BART and random forest principles for improved performance.
• Utilized R and Rcpp for implementation, ensuring accessibility and efficiency.

Main Results:

• BART-BMA demonstrates significant computational efficiency gains over standard BART for datasets with large *p*.
• The algorithm runs in a reasonable time on standard hardware, addressing the "small *n* large *p" challenge.
• Successful application in simulated data and real-world proteomic experiments for disease classification.

Conclusions:

• BART-BMA provides an efficient and effective solution for analyzing high-dimensional data in bioinformatics.
• The method offers a valuable alternative to existing algorithms like BART and random forests.
• Open-source code is available, facilitating further research and application.