230 lines
9.1 KiB
Python
230 lines
9.1 KiB
Python
"""
|
|
Kanevsky all minimum node k cutsets algorithm.
|
|
"""
|
|
import copy
|
|
from collections import defaultdict
|
|
from itertools import combinations
|
|
from operator import itemgetter
|
|
|
|
import networkx as nx
|
|
from .utils import build_auxiliary_node_connectivity
|
|
from networkx.algorithms.flow import (
|
|
build_residual_network,
|
|
edmonds_karp,
|
|
shortest_augmenting_path,
|
|
)
|
|
|
|
default_flow_func = edmonds_karp
|
|
|
|
|
|
__all__ = ["all_node_cuts"]
|
|
|
|
|
|
def all_node_cuts(G, k=None, flow_func=None):
|
|
r"""Returns all minimum k cutsets of an undirected graph G.
|
|
|
|
This implementation is based on Kanevsky's algorithm [1]_ for finding all
|
|
minimum-size node cut-sets of an undirected graph G; ie the set (or sets)
|
|
of nodes of cardinality equal to the node connectivity of G. Thus if
|
|
removed, would break G into two or more connected components.
|
|
|
|
Parameters
|
|
----------
|
|
G : NetworkX graph
|
|
Undirected graph
|
|
|
|
k : Integer
|
|
Node connectivity of the input graph. If k is None, then it is
|
|
computed. Default value: None.
|
|
|
|
flow_func : function
|
|
Function to perform the underlying flow computations. Default value
|
|
edmonds_karp. This function performs better in sparse graphs with
|
|
right tailed degree distributions. shortest_augmenting_path will
|
|
perform better in denser graphs.
|
|
|
|
|
|
Returns
|
|
-------
|
|
cuts : a generator of node cutsets
|
|
Each node cutset has cardinality equal to the node connectivity of
|
|
the input graph.
|
|
|
|
Examples
|
|
--------
|
|
>>> # A two-dimensional grid graph has 4 cutsets of cardinality 2
|
|
>>> G = nx.grid_2d_graph(5, 5)
|
|
>>> cutsets = list(nx.all_node_cuts(G))
|
|
>>> len(cutsets)
|
|
4
|
|
>>> all(2 == len(cutset) for cutset in cutsets)
|
|
True
|
|
>>> nx.node_connectivity(G)
|
|
2
|
|
|
|
Notes
|
|
-----
|
|
This implementation is based on the sequential algorithm for finding all
|
|
minimum-size separating vertex sets in a graph [1]_. The main idea is to
|
|
compute minimum cuts using local maximum flow computations among a set
|
|
of nodes of highest degree and all other non-adjacent nodes in the Graph.
|
|
Once we find a minimum cut, we add an edge between the high degree
|
|
node and the target node of the local maximum flow computation to make
|
|
sure that we will not find that minimum cut again.
|
|
|
|
See also
|
|
--------
|
|
node_connectivity
|
|
edmonds_karp
|
|
shortest_augmenting_path
|
|
|
|
References
|
|
----------
|
|
.. [1] Kanevsky, A. (1993). Finding all minimum-size separating vertex
|
|
sets in a graph. Networks 23(6), 533--541.
|
|
http://onlinelibrary.wiley.com/doi/10.1002/net.3230230604/abstract
|
|
|
|
"""
|
|
if not nx.is_connected(G):
|
|
raise nx.NetworkXError("Input graph is disconnected.")
|
|
|
|
# Address some corner cases first.
|
|
# For complete Graphs
|
|
if nx.density(G) == 1:
|
|
for cut_set in combinations(G, len(G) - 1):
|
|
yield set(cut_set)
|
|
return
|
|
# Initialize data structures.
|
|
# Keep track of the cuts already computed so we do not repeat them.
|
|
seen = []
|
|
# Even-Tarjan reduction is what we call auxiliary digraph
|
|
# for node connectivity.
|
|
H = build_auxiliary_node_connectivity(G)
|
|
H_nodes = H.nodes # for speed
|
|
mapping = H.graph["mapping"]
|
|
# Keep a copy of original predecessors, H will be modified later.
|
|
# Shallow copy is enough.
|
|
original_H_pred = copy.copy(H._pred)
|
|
R = build_residual_network(H, "capacity")
|
|
kwargs = dict(capacity="capacity", residual=R)
|
|
# Define default flow function
|
|
if flow_func is None:
|
|
flow_func = default_flow_func
|
|
if flow_func is shortest_augmenting_path:
|
|
kwargs["two_phase"] = True
|
|
# Begin the actual algorithm
|
|
# step 1: Find node connectivity k of G
|
|
if k is None:
|
|
k = nx.node_connectivity(G, flow_func=flow_func)
|
|
# step 2:
|
|
# Find k nodes with top degree, call it X:
|
|
X = {n for n, d in sorted(G.degree(), key=itemgetter(1), reverse=True)[:k]}
|
|
# Check if X is a k-node-cutset
|
|
if _is_separating_set(G, X):
|
|
seen.append(X)
|
|
yield X
|
|
|
|
for x in X:
|
|
# step 3: Compute local connectivity flow of x with all other
|
|
# non adjacent nodes in G
|
|
non_adjacent = set(G) - X - set(G[x])
|
|
for v in non_adjacent:
|
|
# step 4: compute maximum flow in an Even-Tarjan reduction H of G
|
|
# and step 5: build the associated residual network R
|
|
R = flow_func(H, f"{mapping[x]}B", f"{mapping[v]}A", **kwargs)
|
|
flow_value = R.graph["flow_value"]
|
|
|
|
if flow_value == k:
|
|
# Find the nodes incident to the flow.
|
|
E1 = flowed_edges = [
|
|
(u, w) for (u, w, d) in R.edges(data=True) if d["flow"] != 0
|
|
]
|
|
VE1 = incident_nodes = {n for edge in E1 for n in edge}
|
|
# Remove saturated edges form the residual network.
|
|
# Note that reversed edges are introduced with capacity 0
|
|
# in the residual graph and they need to be removed too.
|
|
saturated_edges = [
|
|
(u, w, d)
|
|
for (u, w, d) in R.edges(data=True)
|
|
if d["capacity"] == d["flow"] or d["capacity"] == 0
|
|
]
|
|
R.remove_edges_from(saturated_edges)
|
|
R_closure = nx.transitive_closure(R)
|
|
# step 6: shrink the strongly connected components of
|
|
# residual flow network R and call it L.
|
|
L = nx.condensation(R)
|
|
cmap = L.graph["mapping"]
|
|
inv_cmap = defaultdict(list)
|
|
for n, scc in cmap.items():
|
|
inv_cmap[scc].append(n)
|
|
# Find the incident nodes in the condensed graph.
|
|
VE1 = {cmap[n] for n in VE1}
|
|
# step 7: Compute all antichains of L;
|
|
# they map to closed sets in H.
|
|
# Any edge in H that links a closed set is part of a cutset.
|
|
for antichain in nx.antichains(L):
|
|
# Only antichains that are subsets of incident nodes counts.
|
|
# Lemma 8 in reference.
|
|
if not set(antichain).issubset(VE1):
|
|
continue
|
|
# Nodes in an antichain of the condensation graph of
|
|
# the residual network map to a closed set of nodes that
|
|
# define a node partition of the auxiliary digraph H
|
|
# through taking all of antichain's predecessors in the
|
|
# transitive closure.
|
|
S = set()
|
|
for scc in antichain:
|
|
S.update(inv_cmap[scc])
|
|
S_ancestors = set()
|
|
for n in S:
|
|
S_ancestors.update(R_closure._pred[n])
|
|
S.update(S_ancestors)
|
|
if f"{mapping[x]}B" not in S or f"{mapping[v]}A" in S:
|
|
continue
|
|
# Find the cutset that links the node partition (S,~S) in H
|
|
cutset = set()
|
|
for u in S:
|
|
cutset.update((u, w) for w in original_H_pred[u] if w not in S)
|
|
# The edges in H that form the cutset are internal edges
|
|
# (ie edges that represent a node of the original graph G)
|
|
if any([H_nodes[u]["id"] != H_nodes[w]["id"] for u, w in cutset]):
|
|
continue
|
|
node_cut = {H_nodes[u]["id"] for u, _ in cutset}
|
|
|
|
if len(node_cut) == k:
|
|
# The cut is invalid if it includes internal edges of
|
|
# end nodes. The other half of Lemma 8 in ref.
|
|
if x in node_cut or v in node_cut:
|
|
continue
|
|
if node_cut not in seen:
|
|
yield node_cut
|
|
seen.append(node_cut)
|
|
|
|
# Add an edge (x, v) to make sure that we do not
|
|
# find this cutset again. This is equivalent
|
|
# of adding the edge in the input graph
|
|
# G.add_edge(x, v) and then regenerate H and R:
|
|
# Add edges to the auxiliary digraph.
|
|
# See build_residual_network for convention we used
|
|
# in residual graphs.
|
|
H.add_edge(f"{mapping[x]}B", f"{mapping[v]}A", capacity=1)
|
|
H.add_edge(f"{mapping[v]}B", f"{mapping[x]}A", capacity=1)
|
|
# Add edges to the residual network.
|
|
R.add_edge(f"{mapping[x]}B", f"{mapping[v]}A", capacity=1)
|
|
R.add_edge(f"{mapping[v]}A", f"{mapping[x]}B", capacity=0)
|
|
R.add_edge(f"{mapping[v]}B", f"{mapping[x]}A", capacity=1)
|
|
R.add_edge(f"{mapping[x]}A", f"{mapping[v]}B", capacity=0)
|
|
|
|
# Add again the saturated edges to reuse the residual network
|
|
R.add_edges_from(saturated_edges)
|
|
|
|
|
|
def _is_separating_set(G, cut):
|
|
"""Assumes that the input graph is connected"""
|
|
if len(cut) == len(G) - 1:
|
|
return True
|
|
|
|
H = nx.restricted_view(G, cut, [])
|
|
if nx.is_connected(H):
|
|
return False
|
|
return True
|