On group-wise ℓp regularization: Theory and efficient algorithms

作者：

Highlights：

• We prove that group-wise ℓp-regularization has algorithmic stability and thus generalizes.

• We derive ADMM and FISTA algorithms for solving group-wise ℓp-regularization.

• We show that for p=5/4, 4/3, 3/2 and 2 the update step has analytical solution.

• We demonstrate that ℓp regularization achieves flexibility in denseness/sparseness modeling.

• We show that group-wise ℓp-regularization has state-of-the-art performance on splice detection.

摘要

Highlights•We prove that group-wise ℓp-regularization has algorithmic stability and thus generalizes.•We derive ADMM and FISTA algorithms for solving group-wise ℓp-regularization.•We show that for p=5/4, 4/3, 3/2 and 2 the update step has analytical solution.•We demonstrate that ℓp regularization achieves flexibility in denseness/sparseness modeling.•We show that group-wise ℓp-regularization has state-of-the-art performance on splice detection.

论文关键词：ℓp Regularization,Convex optimization algorithms,ADMM,FISTA,Algorithmic stability,Lasso,Group Lasso,Bridge regression,Group bridge regression,Splice detection

论文评审过程：Received 16 April 2014, Revised 10 March 2015, Accepted 11 May 2015, Available online 27 May 2015, Version of Record 16 July 2015.

论文官网地址：https://doi.org/10.1016/j.patcog.2015.05.009