# Copyright (C) 2017-2019 by
# Luca Baldesi
# BSD license.
#
# Author: Luca Baldesi (luca.baldesi@unitn.it)
"""Generates graphs with a given eigenvector structure"""
import networkx as nx
from networkx.utils import np_random_state
__all__ = ['spectral_graph_forge']
def _truncate(x):
""" Returns the truncated value of x in the interval [0,1]
"""
if x < 0:
return 0
if x > 1:
return 1
return x
def _mat_spect_approx(A, level, sorteigs=True, reverse=False, absolute=True):
""" Returns the low-rank approximation of the given matrix A
Parameters
----------
A : numpy matrix
level : integer
It represents the fixed rank for the output approximation matrix
sorteigs : boolean
Whether eigenvectors should be sorted according to their associated
eigenvalues before removing the firsts of them
reverse : boolean
Whether eigenvectors list should be reversed before removing the firsts
of them
absolute : boolean
Whether eigenvectors should be sorted considering the absolute values
of the corresponding eigenvalues
Returns
-------
B : numpy matrix
low-rank approximation of A
Notes
-----
Low-rank matrix approximation is about finding a fixed rank matrix close
enough to the input one with respect to a given norm (distance).
In the case of real symmetric input matrix and euclidean distance, the best
low-rank approximation is given by the sum of first eigenvector matrices.
References
----------
.. [1] G. Eckart and G. Young, The approximation of one matrix by another
of lower rank
.. [2] L. Mirsky, Symmetric gauge functions and unitarily invariant norms
"""
import numpy as np
d, V = np.linalg.eigh(A)
d = np.ravel(d)
n = len(d)
if sorteigs:
if absolute:
k = np.argsort(np.abs(d))
else:
k = np.argsort(d)
# ordered from the lowest to the highest
else:
k = range(n)
if not reverse:
k = np.flipud(k)
z = np.zeros((n, 1))
for i in range(level, n):
V[:, k[i]] = z
B = V*np.diag(d)*np.transpose(V)
return B
[docs]@np_random_state(3)
def spectral_graph_forge(G, alpha, transformation='identity', seed=None):
"""Returns a random simple graph with spectrum resembling that of `G`
This algorithm, called Spectral Graph Forge (SGF), computes the
eigenvectors of a given graph adjacency matrix, filters them and
builds a random graph with a similar eigenstructure.
SGF has been proved to be particularly useful for synthesizing
realistic social networks and it can also be used to anonymize
graph sensitive data.
Parameters
----------
G : Graph
alpha : float
Ratio representing the percentage of eigenvectors of G to consider,
values in [0,1].
transformation : string, optional
Represents the intended matrix linear transformation, possible values
are 'identity' and 'modularity'
seed : integer, random_state, or None (default)
Indicator of numpy random number generation state.
See :ref:`Randomness<randomness>`.
Returns
-------
H : Graph
A graph with a similar eigenvector structure of the input one.
Raises
------
NetworkXError
If transformation has a value different from 'identity' or 'modularity'
Notes
-----
Spectral Graph Forge (SGF) generates a random simple graph resembling the
global properties of the given one.
It leverages the low-rank approximation of the associated adjacency matrix
driven by the *alpha* precision parameter.
SGF preserves the number of nodes of the input graph and their ordering.
This way, nodes of output graphs resemble the properties of the input one
and attributes can be directly mapped.
It considers the graph adjacency matrices which can optionally be
transformed to other symmetric real matrices (currently transformation
options include *identity* and *modularity*).
The *modularity* transformation, in the sense of Newman's modularity matrix
allows the focusing on community structure related properties of the graph.
SGF applies a low-rank approximation whose fixed rank is computed from the
ratio *alpha* of the input graph adjacency matrix dimension.
This step performs a filtering on the input eigenvectors similar to the low
pass filtering common in telecommunications.
The filtered values (after truncation) are used as input to a Bernoulli
sampling for constructing a random adjacency matrix.
References
----------
.. [1] L. Baldesi, C. T. Butts, A. Markopoulou, "Spectral Graph Forge:
Graph Generation Targeting Modularity", IEEE Infocom, '18.
https://arxiv.org/abs/1801.01715
.. [2] M. Newman, "Networks: an introduction", Oxford university press,
2010
Examples
--------
>>> import networkx as nx
>>> G = nx.karate_club_graph()
>>> H = nx.spectral_graph_forge(G, 0.3)
>>>
"""
import numpy as np
import scipy.stats as stats
available_transformations = ['identity', 'modularity']
alpha = _truncate(alpha)
A = nx.to_numpy_matrix(G)
n = A.shape[1]
level = int(round(n*alpha))
if transformation not in available_transformations:
msg = '\'{0}\' is not a valid transformation. '.format(transformation)
msg += 'Transformations: {0}'.format(available_transformations)
raise nx.NetworkXError(msg)
K = np.ones((1, n)) * A
B = A
if (transformation == 'modularity'):
B -= np.transpose(K) * K / float(sum(np.ravel(K)))
B = _mat_spect_approx(B, level, sorteigs=True, absolute=True)
if (transformation == 'modularity'):
B += np.transpose(K) * K / float(sum(np.ravel(K)))
B = np.vectorize(_truncate, otypes=[np.float])(B)
np.fill_diagonal(B, np.zeros((1, n)))
for i in range(n-1):
B[i, i+1:] = stats.bernoulli.rvs(B[i, i+1:], random_state=seed)
B[i+1:, i] = np.transpose(B[i, i+1:])
H = nx.from_numpy_matrix(B)
return H
# fixture for pytest
def setup_module(module):
import pytest
numpy = pytest.importorskip('numpy')
scipy = pytest.importorskip('scipy')