Gradient boosting with knockoff filters: a biostatistical approach to variable selection.

Area of Science:

• Statistics
• Machine Learning
• Data Science

Background:

• Increasing data complexity necessitates efficient variable selection methods.
• Controlling False Discovery Rate (FDR) and maintaining statistical power are key challenges.
• Knockoff filters offer a robust approach by creating negative controls for inference.

Purpose of the Study:

• To extend the application of knockoff filters to Light Gradient Boosting Machine (LightGBM).
• To enhance variable selection accuracy and efficiency in the context of big data.
• To improve the interpretability of machine learning models using SHAP values.

Main Methods:

• Integration of knockoff variable generation with LightGBM.
• Utilization of Shapely Additive Explanations (SHAP) for model interpretability.
• Extensive experimentation and simulation studies for validation.

Main Results:

• The proposed method accurately identifies important variables for each class.
• Demonstrated superior performance, speed, and efficiency compared to traditional methods.
• Enhanced interpretability of variable importance through SHAP values.

Conclusions:

• The integration of knockoffs into LightGBM provides a powerful tool for variable selection.
• This approach effectively addresses challenges in big data analysis.
• The method advances statistical modeling and machine learning applications by improving performance and interpretability.