The Cluster Elastic Net for High-Dimensional Regression With Unknown Variable Grouping.

Area of Science:

• Statistics
• Machine Learning
• Bioinformatics

Background:

• High-dimensional regression models often use elastic net for coefficient shrinkage.
• Standard elastic net shrinks all coefficients towards the origin, which can be suboptimal for highly correlated features associated with the response.
• A need exists for regression methods that handle correlated predictors more effectively.

Purpose of the Study:

• To introduce the cluster elastic net, a novel regression method.
• To address the limitations of standard elastic net in scenarios with correlated predictors.
• To develop a method that selectively shrinks coefficients of correlated features towards each other.

Main Methods:

• Proposed the cluster elastic net, which infers feature clusters from data.
• Clustering is based on inter-variable correlation and association with the response variable.
• Employs inferred clusters to guide coefficient shrinkage in regression analysis.

Main Results:

• Demonstrated theoretical advantages of the cluster elastic net over standard methods.
• Simulation studies showed improved performance of the proposed approach.
• Applied the cluster elastic net to HIV drug resistance data, yielding accurate regression.

Conclusions:

• The cluster elastic net offers a more accurate regression approach for high-dimensional data with correlated features.
• Data-driven cluster inference allows for adaptive shrinkage, improving model parsimony and predictive power.
• The method shows promise in applications like bioinformatics and drug resistance analysis.