A New Bayesian Lasso.

Area of Science:

• Statistics
• Computational Statistics
• Machine Learning

Background:

• The Bayesian lasso, introduced by Park and Casella (2008), utilizes scale mixture of normal (SMN) priors.
• Existing methods provide a foundation for Bayesian variable selection in linear models.

Purpose of the Study:

• To propose an alternative Bayesian analysis for the lasso problem.
• To introduce a novel hierarchical formulation for the Bayesian lasso using scale mixture of uniform (SMU) priors.

Main Methods:

• A new hierarchical Bayesian formulation utilizing the scale mixture of uniform (SMU) representation of the Laplace density.
• Development of a Gibbs sampler with tractable full conditional posterior distributions for a fully Bayesian treatment.
• An ECM algorithm for computing Maximum A Posteriori (MAP) estimates.

Main Results:

• The proposed Gibbs sampler exhibits good mixing properties.
• The new Bayesian lasso method demonstrates comparable performance to existing methods in prediction accuracy and variable selection.
• Empirical results and real data analyses validate the effectiveness of the new algorithm.

Conclusions:

• The alternative Bayesian lasso formulation provides a viable and effective approach.
• The developed Gibbs sampler offers computational advantages and robust performance.
• The method shows potential for extension to more general modeling scenarios.