STANDARDIZATION AND THE GROUP LASSO PENALTY.

Area of Science:

• Statistics
• Machine Learning
• Statistical Modeling

Background:

• The original Group Lasso (Yuan & Lin, 2007) penalty was designed for uncorrelated features.
• Its widespread adoption for correlated features may limit effectiveness.
• A need exists for robust group selection methods in the presence of feature correlation.

Purpose of the Study:

• To introduce and evaluate a modified Group Lasso penalty matrix for improved performance with correlated features.
• To demonstrate the statistical underpinnings and practical advantages of the proposed method.
• To extend the method for within-group regularization and provide an efficient fitting algorithm.

Main Methods:

• Development of a "standardized Group Lasso" with a revised penalty matrix.
• Theoretical analysis linking the method to uniformly most powerful invariant tests.
• Empirical validation using real and simulated datasets.
• Extension to a "Ridged Group Lasso" for enhanced within-group regularization.
• Implementation using a group-wise coordinate descent algorithm.

Main Results:

• The standardized Group Lasso demonstrates superior performance over the standard Group Lasso on datasets with correlated features.
• The proposed method shows strong efficacy in both simulated and real-world data analyses.
• The Ridged Group Lasso effectively provides necessary within-group regularization.
• The group-wise coordinate descent algorithm efficiently fits both standardized and Ridged Group Lasso models.

Conclusions:

• The standardized Group Lasso offers a more statistically sound and effective approach for group selection with correlated features.
• This method provides a valuable alternative to the original Group Lasso, enhancing model interpretability and predictive accuracy.
• The developed algorithm facilitates practical application of these advanced group selection techniques.