基于spike-and-slab先验分布的贝叶斯变量选择方法

针对超高维数据,提出一种基于spike-and-slab先验分布的超高维线性回归模型的贝叶斯变量选择方法。该方法继承了弹性网方法和EM算法的优点,以较快的收敛速度来获得稀疏的预测模型。特别地,针对系数的spike-and-slab先验分布设置上,该方法允许系数从不同坐标借力、自动适应已知数据的稀疏信息以及进行多重调整。通过与常用方法的比较,证明了该方法的准确性和有效性。

A Bayesian approach for variable selection using spike-and-slab prior distribution

For ultra-high dimensional data, a Bayesian approach using a novel spike-and-slab prior for variable selection in high-dimensional linear regression models is presented. The proposed method aims to inherit the advantages of the elastic net and the EM algorithm to obtain sparse prediction models with faster convergence speed. Furthermore, a spike-and-slab setting of coefficients which allows for borrowing strength across coordinates, adjust to data sparsity information and exert multiplicity adjustment is proposed. Finally, the accuracy and efficiency of the proposed method are demonstrated via comparisons and analyses with common methods.