The parameter sensitivity of random forests.

Area of Science:

• Computational genomics
• Machine learning in bioinformatics
• Statistical learning applications

Background:

• Random Forest (RF) is a popular supervised machine learning ensemble method.
• Parameter selection is critical for RF model fitting but often overlooked.
• Default RF parameters are frequently used despite limited validation in genomics.

Purpose of the Study:

• To investigate the impact of parameter selection on RF performance in computational genomics.
• To address the gap in understanding RF parameter sensitivity in genomic studies.
• To evaluate the necessity of parameter optimization beyond default settings.

Main Methods:

• Applied the Random Forest algorithm to two distinct biological datasets (low and high p/n ratios).
• Analyzed classification performance and variable importance measures (VIMs) under varying parameter settings.
• Compared performance of optimized parameters against default RF parameters.

Main Results:

• Parameterization significantly impacts RF prediction accuracy and VIMs.
• Optimal parameters differ across datasets with varying variable-to-sample ratios (p/n).
• Parameter optimization substantially improves upon default RF parameter performance.

Conclusions:

• RF parameter performance shows high variability for both low and high p/n genomic data.
• Tuning Random Forest models away from default parameters yields significant performance benefits.
• Parameter optimization is essential for robust RF application in computational genomics.