Thresholding Gini variable importance with a single-trained random forest: An empirical Bayes approach

Area of Science:

• Computational biology
• Statistical learning
• Bioinformatics

Background:

• Random Forests (RFs) are popular for feature selection using Variable Importance Measures (VIMs).
• Current VIM thresholding methods are computationally intensive for high-dimensional data.
• A robust method is needed to control false positives in RF feature selection.

Purpose of the Study:

• To introduce RFlocalfdr, a novel statistical approach for thresholding RF Mean Decrease Impurity (MDI) VIMs.
• To provide a computationally efficient method for feature selection in large datasets.
• To control the false positive rate while identifying significant features.

Main Methods:

• Developed RFlocalfdr based on Efron's empirical Bayes framework.
• Applied RFlocalfdr to threshold MDI importances from RF models.
• Validated using synthetic and real-world health datasets, including large-scale genomics data.

Main Results:

• RFlocalfdr achieves accuracy comparable to existing methods.
• RFlocalfdr is orders of magnitude faster than traditional approaches.
• Successfully thresholded a dataset with 10^6 data points, demonstrating scalability.

Conclusions:

• RFlocalfdr is a computationally efficient and accurate tool for RF feature selection.
• It effectively controls false discovery rates, making it suitable for large-scale biological data.
• Compatible with various RF implementations, enhancing its versatility.