node2vec之小黑尝试

2022-01-10 14:59:07 阅读：190 来源： 互联网

标签：尝试 node node2vec self walk nx alias probs 小黑

参数设定

import warnings
import random
warnings.filterwarnings('ignore')
import argparse
import numpy as np
import networkx as nx
#import node2vec
from gensim.models import Word2Vec
import random
np.random.seed(1)
def parse_args():
    '''
    Parses the node2vec arguments.
    '''
    # 使用parser加载信息
    parser = argparse.ArgumentParser(description="Run node2vec.")
    # 输入文件 
    parser.add_argument('--input', nargs='?', default='../graph/karate.edgelist',
                        help='Input graph path')
    # 输出文件
    parser.add_argument('--output', nargs='?', default='../emb/karate.emb',
                        help='Embeddings path')
    # embedding维度
    parser.add_argument('--dimensions', type=int, default=128,
                        help='Number of dimensions. Default is 128.')
    # 节点序列长度
    parser.add_argument('--walk-length', type=int, default=80,
                        help='Length of walk per source. Default is 80.')
    # 随机游走的次数
    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数量，word2vec参数
    parser.add_argument('--iter', default=10, type=int,
                        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 is 1.')
    # 参数q
    parser.add_argument('--q', type=float, default=1,
                        help='Inout hyperparameter. Default is 1.')
    # 权重
    parser.add_argument('--weighted', dest='weighted', action='store_true',
                        help='Boolean specifying (un)weighted. Default is unweighted.')
    parser.add_argument('--unweighted', dest='unweighted', action='store_false')
    parser.set_defaults(weighted=False)
    # 有向无向
    parser.add_argument('--directed', dest='directed', action='store_true',
                        help='Graph is (un)directed. Default is undirected.')
    parser.add_argument('--undirected', dest='undirected', action='store_false')
    parser.set_defaults(directed=False)
    
    return parser.parse_args(args=[])
    # return parser.parse_known_args()
def read_graph(file):
    graph_x = nx.read_edgelist(file,nodetype = int,create_using = nx.DiGraph())
    for edge in graph_x.edges():
        graph_x[edge[0]][edge[1]]['weight'] = 1
    # 无向操作
    if args.undirected:
        graph_x = graph_x.to_undirected()
    return graph_x
args = parse_args()
nx_G = read_graph(args.input)
nx.draw(nx_G,with_labels = True)
list(nx_G.neighbors(25))

[32, 28, 26]
在这里插入图片描述

采样算法

def alias_setup(probs):
    smaller = []
    larger = []
    Q = np.zeros(len(probs),dtype = int)
    P = np.zeros(len(probs))
    probs = [prob / sum(probs) for prob in probs]
    for t,prob in enumerate(probs):
        P[t] = prob * len(probs)
        if P[t] > 1:
            larger.append(t)
        else:
            smaller.append(t)
    # 准备开始采样
    while larger and smaller:
        large = larger.pop()
        small = smaller.pop()
        Q[small] = large
        P[large] -= (1 - P[small])
        if P[large] < 1:
            smaller.append(large)
        else:
            larger.append(large)
    return Q,P
#J,q = alias_setup([1/2,1/3,1/12,1/12]) 
def alias_draw(J, q):
    length = len(J)
    PP = int(np.floor(np.random.rand() * length))
    QQ = int(np.floor(np.random.rand()))
    if QQ <= q[PP]:
        return PP
    else:
        return J[QQ]

定义模型的Graph类

在这里插入图片描述

class Graph(object):
    def __init__(self,nx_G,is_directed, p, q):
        self.nx_G = nx_G
        self.is_directed = is_directed
        self.p = p
        self.q = q
    def get_alias_edge(self, src, dst):
        nx_G = self.nx_G
        p = self.p
        q = self.q
        unnormalized_probs = []
        for node in sorted(nx_G.neighbors(dst)):
            if node == src:
                prob = nx_G[dst][node]['weight'] / p
            elif nx_G.has_edge(node,src):
                prob = nx_G[dst][node]['weight']
            else:
                prob = nx_G[dst][node]['weight'] / q
            unnormalized_probs.append(prob)
        normalized_probs = [float(prob) / sum(unnormalized_probs) for prob in unnormalized_probs]
        #print(normalized_probs)
        return alias_setup(normalized_probs)
    def preprocess_transition_probs(self):
        is_directed = self.is_directed
        nx_G = self.nx_G
        # 结点 的采样映射
        alias_nodes = {}
        for node in nx_G.nodes():
            unnormalized_probs = [nx_G[node][neighbor]['weight'] for neighbor in sorted(nx_G.neighbors(node))]
            probs = [prob / sum(unnormalized_probs) for prob in unnormalized_probs]
            alias_nodes[node] = alias_setup(probs)
        # 边的采样映射
        alias_edges = {}
        for src,tgt in nx_G.edges():
            unnormalized_probs = self.get_alias_edge(src,tgt)
            alias_edges[(src,tgt)] = unnormalized_probs
            if not is_directed:
                unnormalized_probs = self.get_alias_edge(tgt,src)
                alias_edges[(tgt,src)] = unnormalized_probs
        self.alias_nodes = alias_nodes
        self.alias_edges = alias_edges
        return alias_nodes,alias_edges
    def node2vec_walk(self, walk_length, start_node):
        nx_G = self.nx_G
        alias_nodes = self.alias_nodes
        alias_edges = self.alias_edges
        walk = [start_node]
        while len(walk) < walk_length:
            cur = walk[-1]
            neighbors = sorted(nx_G.neighbors(cur))
            if neighbors:
                if len(walk) == 1:
                	# 只有一个结点的话就对点采样
                    q,p = alias_nodes[cur]
                    sample_index = alias_draw(q,p)
                    walk.append(neighbors[sample_index])
                else:
                	# 超过两个点，就进行边采样
                    pre = walk[-2]
                    q,p = alias_edges[(pre,cur)]
                    sample_index = alias_draw(q,p)
                    walk.append(neighbors[sample_index])
            else:
                break
        return walk
    def simulate_walks(self, num_walks, walk_length):
        nx_G = self.nx_G
        nodes = list(nx_G.nodes())
        walks = []
        for t in range(num_walks):
            random.shuffle(nodes)
            #print('epoch {}'.format(t))
            for node in nodes:
                walks.append(self.node2vec_walk(walk_length,node))
        return walks

word2vec训练部分代码

def learn_embeddings(walks):
    # 将node的类型int转化为string
    walk_lol = []
    for walk in walks:
        tmp = []
        for node in walk:
            tmp.append(str(node))
        walk_lol.append(tmp)
    model = Word2Vec(walk_lol,
                     size = 2,
                     window = args.window_size,
                     min_count = 0,
                     sg = 1,
                     workers = args.workers,
                     iter = args.iter)
    model.wv.save_word2vec_format(args.output)
    return model

聚类

# k-means聚类
from sklearn import  cluster
from sklearn.metrics import adjusted_rand_score
from sklearn.model_selection import train_test_split
import pandas as pd
def draw_cluster(p,q,pos):
    g = Graph(nx_G,False,p,q)
    g.preprocess_transition_probs()
    walks = g.simulate_walks(40,40)
    
    # 训练node2vec模型
    model = learn_embeddings(walks)
    
    # 导入节点名称，获取embedding
    embedding_node=[]
    for i in range(1,35):
        j=str(i)
        embedding_node.append(model[j])
    embedding_node=np.matrix(embedding_node).reshape((34,-1))
    y_pred = cluster.KMeans(n_clusters=4).fit_predict(embedding_node) # 调用 test_RandomForestClassifier
    print(y_pred)
    pos = nx.spring_layout(nx_G)
    nx.draw_networkx_nodes(nx_G,pos,node_color = y_pred,label = True)
    nx.draw_networkx_edges(nx_G,pos,nodelist = nx_G.edges())

实验效果

# p小 q大，偏向宽度优先搜索，模型更具有同构(注意黄色点,起着桥接的属性!!!)
p = 0.1
q = 20
draw_cluster(p,q)

[2 2 3 2 1 1 1 2 3 3 1 2 2 2 0 0 1 2 0 2 0 2 0 0 0 0 0 0 3 0 3 3 0 0]
在这里插入图片描述

# p大 q小，偏向深度优先搜索，模型更具有社群属性
p = 20
q = 0.1
draw_cluster(p,q)

[0 0 3 0 2 2 2 0 3 3 2 0 0 0 1 1 2 0 1 0 1 0 1 1 1 1 1 1 3 1 3 3 1 1]