## 一个脚本打比赛之SMP WEIBO 2016 ## 前言：如何对用户进行精准画像是社交网络分析的基础问题。本文就如何对weibo用户网络提取特征发表一点小的想法，还请尽管拍砖。 数据来源：SMP WEIBO 2016 任务目标：分析用户关联关系与用户发帖内容，通过无监督与有监督方法对用户进行聚类。 ———- 第一部分：筛选source，即判定用户发表的内容是否是垃圾信息。

import matplotlib.pyplot as plt
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
import pandas as pd
import numpy as np
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.decomposition import LatentDirichletAllocation
from time import time
%matplotlib inline

训练数据字段含义： uid: 用户唯一标识，由数字组成 retweet count: 转发数，数字 review count: 评论数，数字 source: 来源，文本 time: 创建时间，时间戳文本(目前有两种格式，yyyy-MM-dd HH:mm:ss和yyyy-MM-dd HH:mm) content: 文本内容（可能包含@信息、表情符信息等）

with open('train/train/train_status.txt','r') as f:lines = f.readlines()
status=[]
for line in lines:status.append(line.strip().split(','))
tr_status = pd.DataFrame(status).loc[:,:5]
tr_status.columns=['uid','retweet','review','source','time','content']
tr_status.to_csv('train_status.csv',index=False)
display(tr_status.head())
with open('valid/valid_status.txt','r') as f:lines = f.readlines()
status=[]
for line in lines:status.append(line.strip().split(','))
v_status = pd.DataFrame(status).loc[:,:5]
v_status.columns=['uid','retweet','review','source','time','content']
v_status.to_csv('valid_status.csv',index=False)
display(v_status.head())

已标注用户字段含义：
uid: 用户唯一标识，由数字组成

gender: 用户性别，m代表男性，f代表女性，None代表此项信息缺失

birthday: 用户出生年份，None代表此项信息缺失

location: 用户地域，部分用户包含省份和城市信息，部分用户只有省份信息，None代表此项信息缺失

with open('train/train/train_labels.txt','r') as f:labels = f.readlines()

labels[0]

‘1832205887||m||1990||四川 None\n’

userHasLabel = [x.split("||")[0].strip() for x in lines]

import pandas as pd
t_labels = pd.read_csv('train/train/train_labels.txt',sep="\\|\\|",header=None)
t_labels.columns = ['uid','gender','birthday','location']v_labels = pd.read_csv('valid/valid_labels.txt',sep="\\|\\|",header=None)
v_labels.columns = ['uid','gender','birthday','location']
print(t_labels.head())
print(v_labels.head())
labeled_nodes = pd.concat([t_labels,v_labels])
labeled_nodes.to_csv('labeled_nodes.csv',index=False,encoding='gbk')

uid gender birthday location 0 1832205887 m 1990 四川 None 1 1737245804 m 1982 吉林 长春 2 2157991124 m 1976 四川 成都 3 2758890931 f 1983 黑龙江 哈尔滨 4 1802646764 m 1981 湖南 长沙 uid gender birthday location 0 1743152063 m 1984 广东 广州 1 1073390982 m 1983 北京 朝阳区 2 2137599524 m 1990 湖北 黄石 3 2279196033 f 1987 江苏 南京 4 1039584863 m 1985 广东 深圳 /home/ll/miniconda3/lib/python3.5/site-packages/ipykernel_launcher.py:2: ParserWarning: Falling back to the ‘python’ engine because the ‘c’ engine does not support regex separators (separators > 1 char and different from ‘\s+’ are interpreted as regex); you can avoid this warning by specifying engine=’python’. /home/ll/miniconda3/lib/python3.5/site-packages/ipykernel_launcher.py:5: ParserWarning: Falling back to the ‘python’ engine because the ‘c’ engine does not support regex separators (separators > 1 char and different from ‘\s+’ are interpreted as regex); you can avoid this warning by specifying engine=’python’. “””

df = pd.concat([tr_status,v_status])
df.loc[:,'uid'] = df['uid'].astype(int)
df = df.merge(labeled_nodes)
display(df.head(10))
display(df.shape)

(331634, 9)

labeled_id = list(t_labels['uid']) + list(v_labels['uid'])

print(len(labeled_id),len(set(labeled_id)))

4467 4467 从第二个用户到最后一个用户均为第一个用户的粉丝 筛选出链接中给出的用户。 @output : nodelist

with open('train/train/train_links.txt','r') as f:t_links = f.readlines()
with open('valid/valid_links.txt','r') as f:v_links = f.readlines()
with open('test/test/test_links.txt','r') as f:te_links = f.readlines()

linklist=[]
for link in t_links:linklist += [str(x) for x in link.strip().split(' ')]
for link in v_links:linklist += [str(x) for x in link.strip().split(' ')]
for link in te_links:linklist += [str(x) for x in link.strip().split(' ')]

print(len(linklist))
print(len(set(linklist)))

721388 308787

nodelist = []
for node in labeled_id:try:linklist.index(str(node))nodelist.append(node)except Exception as e:#print(node)pass
print(len(nodelist))

2476

df = df.set_index('uid').loc[nodelist,:]
display(df.shape)
df.to_csv('labeled_linked_fulltable.csv')

(191119, 8) 至此，将可以筛选的Node筛选出来，具有标签与网络中存在的节点

display(df.shape)
import re 
patt = re.compile('努比亚')
res = filter(patt.match, list(df['source'].drop_duplicates()))
list(res)
# re.findall(patt,list(df['source'].drop_duplicates()))
# list(df['source'].drop_duplicates()).index('努比亚Android')

(191119, 9) [‘努比亚智能手机’]

import pandas as pd
#df= pd.read_csv('labeled_linked_fulltable.csv')
df1 = df[['source','location']]
df1.loc[:,'count'] = 1
diffcount = df1.groupby('source')[['location']].apply(lambda x : x.location.drop_duplicates().count()).to_frame('diff')
count = df1.groupby('source')[['location']].apply(lambda x : x.location.count()).to_frame('count')
source = diffcount.join(count)
display(source.head())

/home/ll/miniconda3/lib/python3.5/site-packages/pandas/core/indexing.py:337: SettingWithCopyWarning: A value is trying to be set on a copy of a slice from a DataFrame. Try using .loc[row_indexer,col_indexer] = value instead See the caveats in the documentation: http://pandas.pydata.org/pandas-docs/stable/indexing.html#indexing-view-versus-copy self.obj[key] = _infer_fill_value(value) /home/ll/miniconda3/lib/python3.5/site-packages/pandas/core/indexing.py:517: SettingWithCopyWarning: A value is trying to be set on a copy of a slice from a DataFrame. Try using .loc[row_indexer,col_indexer] = value instead See the caveats in the documentation: http://pandas.pydata.org/pandas-docs/stable/indexing.html#indexing-view-versus-copy self.obj[item] = s

plt.subplot(211)
source.sort_values('count',ascending=False).head(100)['count'].plot(kind='line',figsize=(30,10),title='count long tail distribution' )
plt.subplot(212)
source.sort_values('count',ascending=False).head(100)['diff'].plot(kind='line',figsize=(30,10))

def hebing(group):
#     display(list(group.content.values))return ' '.join([str(x) for x in list(group.content.values)])
content_merge = df.groupby('source')[['content']].apply(hebing).to_frame('content')
#是否过滤，查看主题变化
content_merge = content_merge.join(source,how='left').reset_index()
#content_merge = content_merge.join(source)[content_merge['count']>10]
display(content_merge.head())tokenizer = lambda s:s.split(' ')
tfv = TfidfVectorizer(tokenizer=tokenizer)
data = tfv.fit_transform(content_merge.content)#词频统计 
#max_df=1, min_df=1
tfc = CountVectorizer(max_features=10000,tokenizer=tokenizer)
tf = tfc.fit_transform(content_merge)

import pickle
with open('tfidf_vec.pkl','wb') as f:pickle.dump(data,f)

from time import time
import pickle
with open('tfidf_vec.pkl','rb') as f:data = pickle.load(f)
#print(data[0])
print('维度约减，自动约减到合适的维度')
print('提取重要特征')
from sklearn.decomposition import NMF, LatentDirichletAllocation
t0 = time()
n_components = 2
nmf = NMF(n_components=2).fit(data)
print("done in %0.3fs." % (time() - t0))

维度约减，自动约减到合适的维度
提取重要特征
done in 1.477s.

with open('nmf_model.pkl','wb') as f:pickle.dump(nmf,f)
print(nmf.reconstruction_err_)
topic = nmf.transform(data)
print(topic)
# if topic.shape[1] ==2 :
#     plt.figure(figsize=(30,30))
#     plt.scatter(x=topic[:,0],y=topic[:,1],c=content_merge['count'])

57.0115343243
[[ 0.02043113  0.        ][ 0.09281784  0.        ][ 0.03913675  0.        ]..., [ 0.05743546  0.03386496][ 0.07525523  0.        ][ 0.01523509  0.        ]]

from sklearn.mixture import GaussianMixture
gmm = GaussianMixture(n_components=2)
gmm.fit(topic)
pred = gmm.predict(topic)
content_merge.loc[:,'pred'] = predfrom sklearn.cluster import KMeans
km = KMeans(n_clusters=3)
km.fit(data)
pred = km.predict(data)
content_merge.loc[:,'pred'] = pred

label = ['荣耀6 Plus','iPhone客户端','虾米音乐移动版','爱相机','华住酒店App','分享按钮']
label = content_merge[content_merge.pred==0].head(100).source
#中文字体显示  
import matplotlib
zhfont = matplotlib.font_manager.FontProperties(fname='/home/ll/.fonts/NotoSansMonoCJKsc-Regular.otf')
# plt.rcParams['font.sans-serif'] = ['Source Han Sans TW', 'sans-serif']
# plt.rc('font', family='Noto Sans Mono CJK SC', size=13)
plt.figure(figsize=(20,20))
plt.scatter(x=topic[:,0],y=topic[:,1],c=pred)
for l in label:pindex = content_merge[content_merge['source']==l].indexplt.annotate(l,xy=(topic[pindex,0],topic[pindex,1]),fontproperties=zhfont)

/home/ll/miniconda3/lib/python3.5/site-packages/matplotlib/font_manager.py:1297: UserWarning: findfont: Font family ['Noto Sans Mono CJK SC'] not found. Falling back to DejaVu Sans(prop.get_family(), self.defaultFamily[fontext]))

说明：图中标注的数据是使用KMEANS算法进行聚类得到的，KMEANS算法使用相似度作为衡量指标，可以看到聚类为0的cluster是用户使用的各种手机平台，包括iphone,huawei等等。
结论：使用TFIDF作为source筛选指标，可以由图看出是较管用的。