Search for the smallest random forest.

Area of Science:

• Computational biology
• Bioinformatics
• Statistical learning

Background:

• Random forests are widely used in high-throughput genomic data analysis.
• Their large size often leads to a "black-box" problem, hindering interpretability.
• Determining the optimal number of trees in a random forest is often subjective.

Purpose of the Study:

• To address the fundamental question of how large a random forest needs to be.
• To develop a method for identifying a small sub-forest that maintains the predictive accuracy of a large random forest.
• To enhance the interpretability of random forests by reducing their complexity.

Main Methods:

• Proposed a novel method to identify representative sub-forests.
• Evaluated the method using extensive simulation studies.
• Validated the approach on a real-world dataset for breast cancer prognosis.

Main Results:

• Identified the existence of small sub-forests (single-digit number of trees) that achieve the prediction accuracy of large random forests (thousands of trees).
• Demonstrated that these sub-forests act as 'representatives' of the entire forest.
• Confirmed that reducing random forest size enhances model interpretability.

Conclusions:

• Sub-forests represent the core predictive power of a random forest.
• It is not necessary to use the entire large random forest for optimal prediction performance.
• Reduced-size random forests are more manageable and less of a 'black-box', improving scientific understanding and application.