码迷,mamicode.com
首页 > 其他好文 > 详细

转载:SVD

时间:2016-11-04 01:14:11      阅读:316      评论:0      收藏:0      [点我收藏+]

标签:通过   并且   技术   alt   min   control   bsp   park   maximum   

ComputeSVD


      
在分布式矩阵有CoordinateMatirx, RowMatrix, IndexedRowMatrix三种。除了CoordinateMatrix之外,IndexedRowMatrixRowMatrix都有computeSVD方法,并且CoordinateMatrixtoIndexedRowMatrix()方法和toRowMatrix()方法可以向IndexedRowMatrix RowMatrix两种矩阵类型转换。
  
因此主要对比 IndexedRowMatrix RowMatrix 两种矩阵类型的 ComputSVD 算法进行分析
   关于SVD内容请参看 维基百科 ,和一篇很棒的博文:《机器学习中的数学》进行了解。

一 算法描述:

           def   computeSVD ( k: Int, computeU: Boolean = false, rCond: Double = 1e-9):         
                       
IndexedRowMatrix  返回类型:  SingularValueDecomposition[IndexedRowMatrix, Matrix]
                        RowMatrix               返回类型:  SingularValueDecomposition[RowMatrix, Matrix] 

                
U                is a RowMatrix of size m x k that satisfies U‘ * U = eye(k),
                
S                  is a Vector of size k, holding the singular values in descending order,
                
V                  is a Matrix of size n x k that satisfies V‘ * V = eye(k).


              
k                 number of leading singular values to keep (0 < k <= n). It might return less than k if there are
                                    numerically zero singular values or there are not enough Ritz values converged before the
                                    maximum number of Arnoldi update iterations is reached.

                
computeU   whether to compute U
                 rCoud         the reciprocal condition number. All singular values smaller than rCond * sigma(0) are treated as zero,
                                    where sigma(0) is the largest singular value.
                 return         SingularValueDecomposition(U, s, V). U = null if computeU = false.

二 选择例子:

构建一个4×5的矩阵M:

      技术分享
矩阵的形式为svdM.txt :
                        1  0  0  0  2
                        0  0  3  0  0
                        0  0  0  0  0
                        0  4  0  0  0

M矩阵的奇异值分解后奇异矩阵s应为:

                        4  0  0  0  0
                           0  3  0  0  0
                           0  0
√5 0  0
                           0  0  0  0  0

我们将通过ComputeSVD函数进行验证.

三 构造矩阵,运行算法并验证结果:   

  <一> 构造RowMatrix矩阵:M
 
        scala> val M = new RowMatrix(sc.textFile("hdfs:///usr/matrix/svdM.txt").map(_.split(‘ ‘))
                                                 .map(_.map(_.toDouble)).map(_.toArray)
                                                 .map(line => Vectors.dense(line)))

        M: org.apache.spark.mllib.linalg.distributed.RowMatrix = org.apache.spark.mllib.linalg.distributed.RowMatrix

 
<二> 调用算法
         scala> val svd = M.computeSVD(4, true)
     
   svd: SingularValueDecomposition[RowMatrix,Matrix]
        
可以看到svd是一个SingularValueDecomposition类型的对像,内部包含一个RowMatrix和一个Matrix用算法,并且此处的RowMatrix就是左奇异向量U,Matrix就是右奇异向量V.


 <三> 验证结果

   SingularValueDecomposition类API如下:
         技术分享


 
矩阵M的左奇异向量U:
        scala> scala> val U = svd.U
                   U: org.apache.spark.mllib.linalg.distributed.RowMatrix = org.apache.spark.mllib.linalg.distributed.RowMatrix
         scala> U.rows.foreach(println)
                    [0.0 ,0.0 ,  -0.9999999999999999 ,  -1.4901161193847656E-8]
                    [0.0 ,1.0 ,0.0 ,0.0]
                    [0.0 ,0.0 ,0.0 ,0.0]
                   [-1.0 ,0.0 ,0.0 ,0.0]


矩阵M的奇异值s:
         scala> val s = svd.s
                   s:  org.apache.spark.mllib.linalg.Vector = [4.0,3.0,2.23606797749979,1.4092648163485167E-8]


矩阵M的右奇异向量V:
         scala> val V = svd.V
                    V: org.apache.spark.mllib.linalg.Matrix =
                    0.0    0.0    -0.44721359549995787     0.8944271909999159
                    -1.0   0.0    0.0    0.0
                    0.0    1.0    0.0    0.0
                    0.0    0.0    0.0    0.0
                    0.0    0.0   -0.8944271909999159       -0.447213595499958


转载:SVD

标签:通过   并且   技术   alt   min   control   bsp   park   maximum   

原文地址:http://www.cnblogs.com/txq157/p/6028686.html

(0)
(0)
   
举报
评论 一句话评论(0
登录后才能评论!
© 2014 mamicode.com 版权所有  联系我们:gaon5@hotmail.com
迷上了代码!