论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

一、本文创新点

1、作者注意到重叠核范数对于张量补全的效果比较好,提出了一种用于张量补全的重叠核范数正则化的非凸变体。

2、该正则化自然引起损失函数的非凸,故提出了一个更有效的优化的算法

3、提出的优化算法有三点优势:一:避免了代价昂贵的张量展开和合并;二:在迭代中保持“稀疏加低秩”结构;三:融入适应的动力。

二、相关工作

1、低秩矩阵学习

                                                                     论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

求解方法:第一步:论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

                 第二步:论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

 

对于求解方程(3),已经有理论保证,理论如下:

                              论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

 

下面来看看具体的例子

                                   论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

求解方法:第一步:论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations(其实就是代入方程2中)

                  第二步:依据上面的步骤,现在应该算论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations的SVD了,但是直接运算代价有点大,所以这里利用了Power method去算

 

2、非凸的低秩正则化

                                         论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

对于非凸正则化,有如下假设

                     论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

常见的非凸正则化有:(1)论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

                                       (2)论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

类似于Lemma 2.1,对于求解带有非凸问题,也有近端梯度步骤

                         论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

对于求解方程(8),也已有理论保证,理论如下:

                               论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

三、非凸低秩张量填充(这里才是本文的重点哈,前面都是些铺垫啦啦啦)

                        论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

直接利用近端梯度算法来求解并不是很乐观,因为论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations并不简单。所以这里先利用近端平均算法(PA算法)

PA算法是解决函数中包含凸但非光滑的函数项,如下:

                                          论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

具体迭代步骤为:

                                    论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

那么对于问题(9),具体的迭代步骤为:

                                       论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

                                   论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

下面利用低秩加稀疏结构,可以避免对张量进行展开和合并操作

                                    论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

                           论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

                     论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

方程(19)和(20)涉及到的计算,下面的定理给出进一步的计算方法

                               论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

                                  论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

                          论文笔记:Efficient Nonconvex Regularized Tensor Completion with Structure-aware Proximal Iterations

四、结论

本文提出了一个非凸正则化的低秩张量补全模型。提出了一种高效的非凸近似平均算法,该算法在迭代过程中保持了“稀疏加低秩”的结构,并引入了自适应动量。收敛到临界点是有保证的。实验结果表明,该算法比已有的方法更有效、更准确。作为未来的工作,利用自动机器学习(AutoML) (Yao et al., 2018b)自适应调整超参数的建议的方法将是有趣的。