冯诺依曼图熵（VNGE）Python实现及近似计算

简介

冯·诺依曼图熵（VNGE）有助于测量图序列中图之间的信息差异和距离。

给定第一个无向图 $G=(V,E,A)$, 其中 $A$ 是对称的邻接矩阵。 度矩阵定义为 $D=diag(d_1,...,d_n)$，则它的拉普拉斯矩阵为 $L=D-A$。其中后者的特征值 ${\lambda }_i$ 被称为拉普拉斯谱。 这里，$H_{vn}(G)$ 为冯诺依曼图熵（von Neumann graph entropy）。

$$H_{vn}(G)=-\sum _{i=1}^n(\frac{{\lambda }_i}{vol(G)}\log \frac{{\lambda }_i}{vol(G)})$$

图的容量为 $vol(G)={\sum }_{i=1}^n{\lambda }_i=trace(L)$。时间复杂度较高，是 $O(n^3)$。

我提供了Python版本代码来计算VNGE。应该注意的是陈等人给出了一种称为FINGER [1]的近似方法，该方法将VNGE的三次复杂度降低为节点和边的数量的线性复杂度。

VNGE和FINGER的代码如下。

代码

# Name: VNGE
# Author: Reacubeth
# Time: 2021/1/25 16:01
# Mail: noverfitting@gmail.com
# Site: www.omegaxyz.com
# *_*coding:utf-8 *_*

import time
import numpy as np
from scipy.sparse.linalg.eigen.arpack import eigsh


def normalized_laplacian(adj_matrix):
    nodes_degree = np.sum(adj_matrix, axis=1)
    nodes_degree_sqrt = 1/np.sqrt(nodes_degree)
    degree_matrix = np.diag(nodes_degree_sqrt)
    eye_matrix = np.eye(adj_matrix.shape[0])
    return eye_matrix - degree_matrix * adj_matrix * degree_matrix


def unnormalized_laplacian(adj_matrix):
    nodes_degree = np.sum(adj_matrix, axis=1)
    degree_matrix = np.diag(nodes_degree)
    return degree_matrix - adj_matrix


def VNGE_exact(adj_matrix):
    start = time.time()
    nodes_degree = np.sum(adj_matrix, axis=1)
    c = 1.0 / np.sum(nodes_degree)
    laplacian_matrix = c * unnormalized_laplacian(adj_matrix)
    eigenvalues, _ = np.linalg.eig(laplacian_matrix)
    eigenvalues[eigenvalues < 0] = 0
    pos = eigenvalues > 0
    H_vn = - np.sum(eigenvalues[pos] * np.log2(eigenvalues[pos]))
    print('H_vn exact:', H_vn)
    print('Time:', time.time() - start)


def VNGE_FINGER(adj_matrix):
    start = time.time()
    nodes_degree = np.sum(adj_matrix, axis=1)
    c = 1.0 / np.sum(nodes_degree)
    edge_weights = 1.0 * adj_matrix[np.nonzero(adj_matrix)]
    approx = 1.0 - np.square(c) * (np.sum(np.square(nodes_degree)) + np.sum(np.square(edge_weights)))
    laplacian_matrix = unnormalized_laplacian(adj_matrix)
    '''
    eigenvalues, _ = np.linalg.eig(laplacian_matrix)
    eig_max = c * max(eigenvalues)
    '''
    eig_max, _ = eigsh(laplacian_matrix, 1, which='LM')
    eig_max = eig_max[0] * c
    H_vn = - approx * np.log2(eig_max)
    print('H_vn approx:', H_vn)
    print('Time:', time.time() - start)


nodes_num = 3000
sparsity = 0.01


tmp_m = np.random.uniform(0, 1, (nodes_num, nodes_num))
pos1 = tmp_m > sparsity
pos2 = tmp_m <= sparsity
tmp_m[pos1] = 0
tmp_m[pos2] = 1
tmp_m = np.triu(tmp_m)
adj_m = tmp_m + np.transpose(tmp_m)

VNGE_exact(adj_m)
VNGE_FINGER(adj_m)

结果

H_vn exact: 11.496599468152386
Time: 13.172455072402954
H_vn approx: 10.63029083591871
Time: 0.23734617233276367

[1] Chen, Pin-Yu, et al. “Fast incremental von neumann graph entropy computation: Theory, algorithm, and applications.” International Conference on Machine Learning. PMLR, 2019.

更多内容访问 omegaxyz.com
网站所有代码采用Apache 2.0授权
网站文章采用知识共享许可协议BY-NC-SA4.0授权
© 2021 • OmegaXYZ-版权所有 转载请注明出处