Variable importance-weighted Random Forests.

Area of Science:

• Bioinformatics
• Machine Learning in Biology

Background:

• Random Forests (RF) is a widely used classification and regression algorithm in biological studies.
• RF performance declines with a high number of features, leading to methods like feature elimination RF.
• Existing methods have limitations in rigid feature selection and increased tree correlations.

Purpose of the Study:

• To propose a novel Random Forests approach that enhances prediction accuracy.
• To address the performance degradation of standard RF with high dimensionality.
• To improve feature selection by incorporating variable importance scores.

Main Methods:

• Introduced variable importance-weighted Random Forests (viRF).
• viRF samples features based on their importance scores at each node, unlike standard RF's equal probability sampling.
• The best split is selected from these importance-weighted sampled features.

Main Results:

• viRF demonstrated improved performance over standard RF and feature elimination RF in simulations and real-world data.
• The method showed enhanced accuracy in both classification and regression tasks.
• Performance gains were observed particularly in the presence of weak signals and high noise.

Conclusions:

• Incorporating variable importance into feature selection improves utilization of informative features.
• viRF offers better prediction accuracy by balancing informative and less informative features.
• An R package 'viRandomForests' is available for public use.