动手复现Node2Vec代码并实现可视化分析
动手复现Node2Vec代码并实现可视化分析
- 环境配置
- 用node2vec现成工具包实现分析
-
- 导入工具包
- 导入悲惨世界人物数据集
- 可视化
- 构建node2vec模型
- 节点embedding聚类可视化
- 对edge做embedding
- 动手实现node2vec (核心：alias sampling算法)
-
- 导入工具包
- 输入基本参数信息
- 载入图
- Alias Sampling
- 生成随机游走序列
- 采样得到所有随机游走序列
- 利用word2vec训练node2vec
- 可视化分析
- 查找相似节点
- 聚类分析
原文链接：node2vec: Scalable Feature Learning for Networks
论文精读：node2vec，站在DeepWalk巨人肩膀上再进一步
手动防爬虫，作者CSDN：总是重复名字我很烦啊，联系邮箱daledeng123@163.com
环境配置
| 工具包 | 作用 |
|---|---|
| networkx | 图网络可视化分析 |
| pandas | 数据分析 |
| numpy | 数据分析 |
| node2vec | 现成的node2vec直接调用 |
| gensim | 自然语言处理相关工具(word2vec) |
| scikit-learn | 机器学习(KMeans聚类) |
| argparse | 默认参数汇总 |
| random | 生成随机数和打乱 |
!pip install networkx pandas numpy node2vec gensim scikit-learn -i https://pypi.tuna.tsinghua.edu.cn/simple
用node2vec现成工具包实现分析
导入工具包
import networkx as nx
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings('ignore')
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False
导入悲惨世界人物数据集
G = nx.les_miserables_graph()
可视化
plt.figure(figsize=(15,14))
pos = nx.spring_layout(G, seed=5)
nx.draw(G, pos, with_labels=True)
plt.show()
构建node2vec模型
from node2vec import Node2Vec
node2vec = Node2Vec(G,
dimensions=32, #嵌入维度
p=2, #return参数
q=0.5, #out参数
walk_length=30, #随机游走最大长度
num_walks=600, #每个节点作为起始节点的随机游走个数
workers=4) #并行线程数
# p=1, q=0.5, n_clusters=6, DFS深度优先
# p=1, q=2, n_clusters=3, BFS广度优先
model = node2vec.fit(window=3, #skpi-gram窗口大小
min_count=1, #设置阈值忽略低频节点
batch_words=4, #每个线程处理的数据量
)
X = model.wv.vectors
节点embedding聚类可视化
from sklearn.cluster import KMeans
cluster_labels = KMeans(n_clusters=3).fit(X).labels_
# 将networkx中的节点和词向量中的节点对应
colors = []
nodes = list(G.nodes)
for node in nodes:
# 获取这个节点在embedding中的索引号
idx = model.wv.key_to_index[str(node)]
colors.append(cluster_labels[idx])
# 可视化
plt.figure(figsize=(15,14))
pos = nx.spring_layout(G, seed=10)
nx.draw(G, pos, node_color=colors, with_labels=True)
对edge做embedding
from node2vec.edges import HadamardEmbedder
edges_embs = HadamardEmbedder(keyed_vectors=model.wv)
# 计算所有edge的embedding
edges_kv = edges_embs.as_keyed_vectors()
# 查看最相似的节点对
edges_kv.most_similar(str(('Bossuet', 'Valjean')))
>>[("('Bahorel', 'Valjean')", 0.9560104012489319),
("('Prouvaire', 'Valjean')", 0.926855742931366),
("('Courfeyrac', 'Valjean')", 0.9163148403167725),
("('Enjolras', 'Valjean')", 0.900037944316864),
("('Combeferre', 'Valjean')", 0.8826196789741516),
("('Feuilly', 'Valjean')", 0.8544862866401672),
("('Joly', 'Valjean')", 0.8330731391906738),
("('MmeHucheloup', 'Valjean')", 0.8265642523765564),
("('Grantaire', 'Valjean')", 0.7791286706924438),
("('Enjolras', 'Javert')", 0.7411160469055176)]
动手实现node2vec (核心：alias sampling算法)
导入工具包
import networkx as nx
import pandas as pd
import numpy as np
from gensim.models import Word2Vec
import random
import argparse
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings('ignore')
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False
输入基本参数信息
def parse_args():
# 使用parser加载信息
parser = argparse.ArgumentParser(description='Run node2vec.')
# 输入文件：邻接表
parser.add_argument('--input', nargs='?', default='karate.edgelist', help='Input graph path')
# 输出文件：节点嵌入表
parser.add_argument('--output', nargs='?', default='karate.emb', help='Embedding path')
# embedding嵌入向量维度
parser.add_argument('--dimensions', type=int, default=128, help='Number of dimensions. Default is 128.')
# 随机游走序列长度
parser.add_argument('--walk-length', type=int, default=80, help='Length of walk per source. Default is 80.')
# 每个节点生成随机游走序列次数
parser.add_argument('--num-walks', type=int, default=10, help='Number of walks per source. Default is 10.')
# word2vec窗口大小，word2vec参数
parser.add_argument('--window-size', type=int, default=10, help='Context size for optimization. Default is 10.')
# SGD优化时epoch输了，word2vec参数
parser.add_argument('--iter', type=int, default=1, help='Number of epochs in SGD')
# 并行化核数，word2vec参数
parser.add_argument('--workers', type=int, default=8, help='Number of parallel workers. Default is 8.')
# 参数p
parser.add_argument('--p', type=float, default=1, help='Return hyperparameter. Default os 1.')
# 参数q
parser.add_argument('--q', type=float, default=1, help='Inout hyperparameter. Default os 1.')
# 连接是否带权重
parser.add_argument('--weight', dest='weight', action='store_true', help='Boolean specifying (un)weighted. Default is unweighted.')
parser.add_argument('--unweight', dest='unweight', action='store_false')
parser.set_defaults(weighted=False)
# 有向图还是无向图
parser.add_argument('--directed', dest='directed', action='store_true', help='Graph is (un)directd. Default is undirected,')
parser.add_argument('--undirected', dest='undirected', action='store_false')
parser.set_defaults(weighted=False)
return parser.parse_args(args=[])
args = parse_args()
载入图
# 连接带权重
if args.weighted:
G = nx.read_edgelist(args.input, nodetype =int, data=(('weight', float),), create_using=nx.DiGraph())
# 连接不带权重
else:
G = nx.read_edgelist(args.input, nodetype=int, create_using=nx.DiGraph())
for edge in G.edges():
G[edge[0]][edge[1]]['weight'] = np.abs(np.random.randn())
# 无向图
if not args.directed:
G = G.to_undirected()
pos = nx.spring_layout(G, seed=4)
nx.draw(G, pos, with_labels=True)
Alias Sampling
def alias_setup(probs):
K = len(probs)
q = np.zeros(K)
J = np.zeros(K, dtype=np.int8)
smaller = []
larger = []
# 将各个概率分为两组，一组的概率值大于1，另一组的概率小于1
for kk, prob in enumerate(probs):
q[kk] = K * prob # 每类事件的概率乘以事件个数
# 判断多了还是少了
if q[kk] < 1.0:
smaller.append(kk)
else:
larger.append(kk)
# 使用贪心算法，将概率值小于1的不断填满
while len(smaller) > 0 and len(larger) > 0:
small = smaller.pop()
large = larger.pop()
J[small] = large
# 更新概率值，劫富济贫，削峰填谷
q[large] = q[large] - (1 - q[small])
if q[large] < 1.0:
smaller.append(large) # 把被削的土地给穷人
else:
larger.append(large)
return J, q
# 执行O（1）采样
def alias_draw(J, q):
K = len(J)
kk = int(np.floor(np.random.rand()*K)) # 生成一个1-k的随机整数
if np.random.rand() < q[kk]:
return kk # 取自己本身事件
else:
return J[kk] # 取alias事件
def get_alias_edge(src, dst):
p = args.p
q = args.q
unnormalized_probs = []
# 论文3.2.2核心算法，计算各条边的转移权重
for dst_nbr in sorted(G.neighbors(dst)):
if dst_nbr == src:
unnormalized_probs.append(G[dst][dst_nbr]['weight']/p)
elif G.has_edge(dst_nbr, src):
unnormalized_probs.append(G[dst][dst_nbr]['weight'])
else:
unnormalized_probs.append(G[dst][dst_nbr]['weight']/q)
# 归一化
norm_const = sum(unnormalized_probs)
normalized_probs = [float(u_prob)/norm_const for u_prob in unnormalized_probs]
return alias_setup(normalized_probs)
is_directed = args.directed
alias_nodes = {}
# 节点概率alias sampling归一化
for node in G.nodes():
unnormalized_probs = [G[node][nbr]['weight'] for nbr in sorted(G.neighbors(node))]
norm_const = sum(unnormalized_probs)
normalized_probs = [float(u_prob)/norm_const for u_prob in unnormalized_probs]
alias_nodes[node] = alias_setup(normalized_probs)
if node==25:
print('25号节点')
print(unnormalized_probs)
print(norm_const)
print(normalized_probs)
print(alias_nodes[node])
>>25号节点
[1.2598288084081546, 1.1050425433820839, 0.2577415056206481]
2.622612857410887
[0.48037162818300644, 0.42135176004323077, 0.09827661177376266]
(array([0, 0, 1], dtype=int8), array([1. , 0.55888512, 0.29482984]))
alias_edges = {}
triads = {}
# 边概率alias sampling和归一化
if is_directed:
for edge in G.edges():
alias_edges[edge] = get_alias_edge(edge[0], edge[1])
else:
for edge in G.edges():
alias_edges[edge] = get_alias_edge(edge[0], edge[1])
alias_edges[(edge[1], edge[0])] = get_alias_edge(edge[1], edge[0])
生成随机游走序列
def node2vec_walk(walk_length, start_node):
# 上一步计算出的alias table，完成O(1)采样
walk = [start_node]
# 知道生成长度为walk_length的节点序列未知
while len(walk) < walk_length:
cur = walk[-1]
# 对邻居节点排序，目的是和alias table计算时的顺序对应起来
cur_nbrs = sorted(G.neighbors(cur))
if len(cur_nbrs) > 0:
# 节点序列只有一个节点的情况
if len(walk) == 1:
walk.append(cur_nbrs[alias_draw(alias_nodes[cur][0], alias_nodes[cur][1])])
# 节点序列大于一个节点的情况
else:
# 看前一个节点
prev = walk[-2]
next = cur_nbrs[alias_draw(alias_edges[(prev, cur)][0], alias_edges[(prev, cur)][1])]
walk.append(next)
else:
break
return walk
采样得到所有随机游走序列
def simulate_walks(num_walks, walk_length):
walks = []
nodes = list(G.nodes())
print('Walk iteration')
for walk_iter in range(num_walks):
print(str(walk_iter+1), '/', str(num_walks))
# 打乱顺序
random.shuffle(nodes)
for node in nodes:
walks.append(node2vec_walk(walk_length=walk_length, start_node=node))
return walks
walks = simulate_walks(args.num_walks, args.walk_length)
# 把node的类型int转化为str
walk_str = []
for walk in walks:
tmp = []
for node in walk:
tmp.append(str(node))
walk_str.append(tmp)
利用word2vec训练node2vec
model = Word2Vec(walk_str, vector_size=args.dimensions, window=args.window_size, min_count=0, sg=1, workers=args.workers)
model.wv.save_word2vec_format(args.output)
可视化分析
model.wv.get_vector('17')
>>array([ 0.18560958, 0.04532452, 0.26840374, 0.03078318, 0.02398106,
-0.09106599, 0.17842552, 0.06163856, 0.0530385 , -0.14195229,
0.04106649, -0.1773398 , -0.03475689, -0.15610787, 0.07816559,
-0.06591983, 0.02470083, 0.02963566, -0.16297406, 0.06844087,
0.21037795, 0.15360692, 0.11143462, -0.0366699 , -0.29888242,
0.17556562, -0.04650875, 0.12099634, 0.18204564, -0.06231313,
-0.3934941 , 0.16333763, -0.13903633, -0.06437788, -0.13468122,
0.11245078, 0.29344803, 0.27591076, 0.1982137 , -0.08944517,
-0.05133532, 0.16017602, -0.02978732, -0.06244398, 0.16406792,
0.09379737, 0.00436372, -0.23881316, -0.11517377, 0.12249801,
0.01371407, 0.0331029 , 0.23663211, 0.17176941, 0.10793246,
-0.04616706, 0.19773844, 0.03704267, -0.14675048, -0.2599338 ,
0.08897672, -0.19225457, 0.09435391, -0.0571735 , 0.04452999,
-0.12241848, -0.01499153, 0.07220587, -0.15018323, -0.08562861,
0.18255252, 0.06718722, -0.22175026, 0.11272267, 0.03694518,
0.01602137, -0.05840521, -0.11858806, -0.00055711, 0.1841024 ,
0.12401607, -0.18123557, 0.10295101, 0.11569355, 0.07272708,
-0.02486938, 0.23009172, -0.09448099, 0.26088616, 0.08620758,
-0.10338984, -0.02792386, 0.00071772, 0.05634106, 0.06651364,
0.15144321, -0.18755764, -0.21857035, -0.02649569, -0.1717046 ,
-0.16789995, 0.01954506, 0.12292644, 0.05823366, 0.10581642,
-0.11863767, 0.09803073, -0.15999381, -0.03844869, -0.00235334,
0.00488694, 0.05818262, 0.0406672 , 0.06263107, 0.06486863,
-0.01384014, 0.15145428, 0.22274497, 0.15133294, 0.14930773,
0.14660034, -0.01708281, 0.06902762, 0.0583788 , 0.07930274,
-0.23277247, 0.0493211 , 0.09819821], dtype=float32)
查找相似节点
print(model.wv.most_similar('25'))
>>print(model.wv.most_similar('25'))
print(model.wv.most_similar('25'))
[('26', 0.9905608296394348), ('24', 0.974091649055481), ('28', 0.9634271264076233), ('30', 0.8659281730651855), ('32', 0.8056477904319763), ('27', 0.7184703946113586), ('33', 0.6937069892883301), ('15', 0.6862101554870605), ('23', 0.6689950823783875), ('21', 0.6680737137794495)]
聚类分析
from sklearn.cluster import KMeans
X = model.wv.vectors
cluster_labels = KMeans(n_clusters=3, random_state=9).fit(X).labels_
colors = []
nodes = list(G.nodes)
for node in nodes:
idx = model.wv.key_to_index[str(node)]
colors.append(cluster_labels[idx])
pos = nx.spring_layout(G, seed=4)
nx.draw(G, pos, node_color = colors, with_labels=True)
