我有一个Kosaraju算法的实现，可以在Python中找到SCC。下面的代码包含一个递归版本（适合小型测试用例）和一个非递归版本（由于实际数据集的大小，我最终需要使用它）。
我已经在一些测试数据集中运行了递归和非递归版本，并获得了正确的答案。但是，在我最终需要使用的更大数据集上运行它会产生错误的结果。遍历真实数据并不是真正的选择，因为它包含将近一百万个节点。
我的问题是我不知道如何从这里继续。我的怀疑是我要么在测试用例中忘记了某个图形星座图案例，要么对这种算法的工作原理有更根本的误解。
#!/usr/bin/env python3
import heapq
class Node():
"""A class to represent nodes in a DirectedGraph. It has attributes for
performing DFS."""
def __init__(self, i):
self.id = i
self.edges = []
self.rev_edges = []
self.explored = False
self.fin_time = 0
self.leader = 0
def add_edge(self, edge_id):
self.edges.append(edge_id)
def add_rev_edge(self, edge_id):
self.rev_edges.append(edge_id)
def mark_explored(self):
self.explored = True
def set_leader(self, leader_id):
self.leader = leader_id
def set_fin_time(self, fin_time):
self.fin_time = fin_time
class DirectedGraph():
"""A class to represent directed graphs via the adjacency list approach.
Each dictionary entry is a Node."""
def __init__(self, length, list_of_edges):
self.nodes = {}
self.nodes_by_fin_time = {}
self.length = length
self.fin_time = 1 # counter for the finishing time
self.leader_count = 0 # counter for the size of leader nodes
self.scc_heapq = [] # heapq to store the ssc by size
self.sccs_computed = False
for n in range(1, length + 1):
self.nodes[str(n)] = Node(str(n))
for n in list_of_edges:
ns = n[0].split(' ')
self.nodes[ns[0]].add_edge(ns[1])
self.nodes[ns[1]].add_rev_edge(ns[0])
def n_largest_sccs(self, n):
if not self.sccs_computed:
self.compute_sccs()
return heapq.nlargest(n, self.scc_heapq)
def compute_sccs(self):
"""First compute the finishing times and the resulting order of nodes
via a DFS loop. Second use that new order to compute the SCCs and order
them by their size."""
# Go through the given graph in reverse order, computing the finishing
# times of each node, and create a second graph that uses the finishing
# times as the IDs.
i = self.length
while i > 0:
node = self.nodes[str(i)]
if not node.explored:
self.dfs_fin_times(str(i))
i -= 1
# Populate the edges of the nodes_by_fin_time
for n in self.nodes.values():
for e in n.edges:
e_head_fin_time = self.nodes[e].fin_time
self.nodes_by_fin_time[n.fin_time].add_edge(e_head_fin_time)
# Use the nodes ordered by finishing times to calculate the SCCs.
i = self.length
while i > 0:
self.leader_count = 0
node = self.nodes_by_fin_time[str(i)]
if not node.explored:
self.dfs_leaders(str(i))
heapq.heappush(self.scc_heapq, (self.leader_count, node.id))
i -= 1
self.sccs_computed = True
def dfs_fin_times(self, start_node_id):
stack = [self.nodes[start_node_id]]
# Perform depth-first search along the reversed edges of a directed
# graph. While doing this populate the finishing times of the nodes
# and create a new graph from those nodes that uses the finishing times
# for indexing instead of the original IDs.
while len(stack) > 0:
curr_node = stack[-1]
explored_rev_edges = 0
curr_node.mark_explored()
for e in curr_node.rev_edges:
rev_edge_head = self.nodes[e]
# If the head of the rev_edge has already been explored, ignore
if rev_edge_head.explored:
explored_rev_edges += 1
continue
else:
stack.append(rev_edge_head)
# If the current node has no valid, unexplored outgoing reverse
# edges, pop it from the stack, populate the fin time, and add it
# to the new graph.
if len(curr_node.rev_edges) - explored_rev_edges == 0:
sink_node = stack.pop()
# The fin time is 0 if that node has not received a fin time.
# Prevents dealing with the same node twice here.
if sink_node and sink_node.fin_time == 0:
sink_node.set_fin_time(str(self.fin_time))
self.nodes_by_fin_time[str(self.fin_time)] = \
Node(str(self.fin_time))
self.fin_time += 1
def dfs_leaders(self, start_node_id):
stack = [self.nodes_by_fin_time[start_node_id]]
while len(stack) > 0:
curr_node = stack.pop()
curr_node.mark_explored()
self.leader_count += 1
for e in curr_node.edges:
if not self.nodes_by_fin_time[e].explored:
stack.append(self.nodes_by_fin_time[e])
###### Recursive verions below ###################################
def dfs_fin_times_rec(self, start_node_id):
curr_node = self.nodes[start_node_id]
curr_node.mark_explored()
for e in curr_node.rev_edges:
if not self.nodes[e].explored:
self.dfs_fin_times_rec(e)
curr_node.set_fin_time(str(self.fin_time))
self.nodes_by_fin_time[str(self.fin_time)] = Node(str(self.fin_time))
self.fin_time += 1
def dfs_leaders_rec(self, start_node_id):
curr_node = self.nodes_by_fin_time[start_node_id]
curr_node.mark_explored()
for e in curr_node.edges:
if not self.nodes_by_fin_time[e].explored:
self.dfs_leaders_rec(e)
self.leader_count += 1
要运行：
#!/usr/bin/env python3
import utils
from graphs import scc_computation
# data = utils.load_tab_delimited_file('data/SCC.txt')
data = utils.load_tab_delimited_file('data/SCC_5.txt')
# g = scc_computation.DirectedGraph(875714, data)
g = scc_computation.DirectedGraph(11, data)
g.compute_sccs()
# for e, v in g.nodes.items():
# print(e, v.fin_time)
# for e, v in g.nodes_by_fin_time.items():
# print(e, v.edges)
print(g.n_largest_sccs(20))
最复杂的测试用例（SCC_5.txt）：
1 5
1 4
2 3
2 11
2 6
3 7
4 2
4 8
4 10
5 7
5 5
5 3
6 8
6 11
7 9
8 2
8 8
9 3
10 1
11 9
11 6
该测试用例的绘制：https://imgur.com/a/LA3ObpN
这会产生4个SCC：
答案 0 :(得分：0)
好的，我找出了丢失的案件。该算法在连接非常紧密的图和重复的边上无法正确执行。这是我上面发布的测试用例的调整版本，具有重复的边缘和更多的边缘，可将整个图形变成一个大的SCC。
1 5
1 4
2 3
2 6
2 11
3 2
3 7
4 2
4 8
4 10
5 1
5 3
5 5
5 7
6 8
7 9
8 2
8 2
8 4
8 8
9 3
10 1
11 9
11 6