Surrogate minimal depth as an importance measure for variables in random forests

Area of Science:

• Bioinformatics
• Machine Learning
• Genomics

Background:

• Random forest is effective for omics data analysis, including classification, regression, and variable selection.
• Interpreting variable importance in complex, high-dimensional omics data is challenging due to intricate relationships between predictors.

Purpose of the Study:

• To introduce a novel variable selection approach, Surrogate Minimal Depth (SMD).
• To enhance the interpretability of variable importance in high-dimensional omics data by considering variable relationships.

Main Methods:

• Developed Surrogate Minimal Depth (SMD) by integrating surrogate variables into the Minimal Depth (MD) concept.
• Applied SMD to simulated correlation patterns to assess its ability to reconstruct relationships and improve variable selection.

Main Results:

• SMD successfully reconstructed simulated correlation patterns, demonstrating its capability to account for variable interdependencies.
• SMD exhibited higher empirical power in identifying causal variables compared to state-of-the-art methods and MD, with comparable stability.
• The method provides improved insights into the complex interplay between predictor variables and outcomes.

Conclusions:

• Surrogate Minimal Depth (SMD) is a promising advancement for variable selection in high-dimensional omics data.
• SMD offers a more nuanced understanding of predictor variable relationships and their impact on outcomes.
• The approach facilitates more accurate identification of causal variables in complex datasets.