基于Capped L1罚函数的组稀疏模型

The Sparse Group Model Based on Capped L1 Penalty

近年来,稀疏化研究在人工智能领域非常流行。变量之间往往存在组结构,Group Lasso利用这种组结构并且可以实现变量组选择。将Capped L1罚推广到变量组选择情形下,提出能够和Group Lasso一样实现变量组选择的GroupCapped L1,然后利用块坐标下降算法求解Group Capped L1的最优化问题。通过实验证明与Group Lasso相比,所提出的Group Capped L1是一种有竞争力的方法。

The research of sparsity is quite hot in the field of artificial intelligence in recent years.The group structure which exists in variables is commonly seen in many settings.The Group Lasso uses the group structure and performs group selection.Extends the Capped L1 penalty to the group selection settings,and proposes the Group Capped L1which performs group selection like the Group Lasso.Uses the block co.ordinate descent algorithm to solve the optimization of the Group Capped L1.The experiments show that the Group Capped L1 proposed is a competitive method when compared with the Group Lasso.