CUR分解
要理解CUR分解,需要先看下SVD分解。SVD理论以及Python实现
算法流程
给定输入的矩阵A。
A = C ∗ U ∗ R A = C* U *R A=C∗U∗R
- 随机选r个列构成C和r个行构成R(也可以使用,平方和加权过的行和列(常用))
- 然后选取W矩阵(C和R的交集,也就是被选出来的部分,在C和R中同时出现的A矩阵中的位置。)
- 对W做SVD分解,得到 X ∑ Y T X\sum Y^T X∑YT
- 对 ∑ \sum ∑做广义逆矩阵 ( ∑ ) + (\sum)^+ (∑)+,也就是只有非0元的部分才变成原来的倒数。
- U = Y ∗ ( ∑ ) + ∗ X T U = Y*(\sum)^+* X^T U=Y∗(∑)+∗XT
Python实现
- 导入包
import numpy as np
- 数据
A = np.linspace(0, 14, 15).reshape((3, -1))
- 算法
def CUR(A, n):
A_sq = A ** 2
sum_A_sq = np.sum(A_sq)
sum_A_sq_0 = np.sum(A_sq, axis=0)
sum_A_sq_1 = np.sum(A_sq, axis=1)
P_x_c = sum_A_sq_0 / sum_A_sq
P_x_r = sum_A_sq_1 / sum_A_sq
r, c = A.shape
c_index = [np.random.choice(np.arange(0, c), p=P_x_c) for i in range(n)]
r_index = [np.random.choice(np.arange(0, r), p=P_x_r) for i in range(n)]
# print(c_index, r_index)
C = A[:, c_index]
R = A[r_index, :]
W = C[r_index]
# print(C, R, W)
def SVD(A, n):
M = np.dot(A, A.T)
eigval, eigvec = np.linalg.eig(M)
indexes = np.argsort(-eigval)[:n]
U = eigvec[:, indexes]
sigma_sq = eigval[indexes]
M = np.dot(A.T, A)
eigval, eigvec = np.linalg.eig(M)
indexes = np.argsort(-eigval)[:n]
V = eigvec[:, indexes]
sigma = sigma_sq # not diag and not sqrt
return U, sigma, V
X, sigma, Y = SVD(W, n)
for i in range(len(sigma)):
if sigma[i] == 0:
continue
else:
sigma[i] = 1 / sigma[i]
sigma = np.diag(sigma)
U = np.dot(np.dot(Y, sigma), X.T)
return np.dot(np.dot(C, U), R)
- 调用
CUR(A, 3)