第一部分
代码一:python访问数据库
import pandas as pd
from sqlalchemy import create_engine
engine = create_engine('mysql+pymysql://root:102011@localhost/test?charset=utf8')
sql = pd.read_sql('all_gzdata', engine, chunksize = 10000)
'''
用create_engine建立连接,连接地址的意思依次为“数据库格式(mysql)+程序名(pymysql)+账号密码@地址端口/数据库名(test)”,最后指定编码为utf8;
all_gzdata是表名,engine是连接数据的引擎,chunksize指定每次读取1万条记录。这时候sql是一个容器,未真正读取数据。
'''
代码二:网页类型统计
counts = [ i['fullURLId'].value_counts() for i in sql] #按次10000存取,逐块统计
counts = counts.copy()
counts = pd.concat(counts).groupby(level=0).sum() #合并统计结果,把相同的统计项合并(即按index分组并求和)
counts = counts.reset_index() #重新设置index,将原来的index作为counts的一列。
counts.columns = ['index', 'num'] #重新设置列名,主要是第二列,默认为0
counts['type'] = counts['index'].str.extract('(\d{3})') #提取前三个数字作为类别id
counts_ = counts[['type', 'num']].groupby('type').sum() #按类别合并
counts_.sort_values('num', ascending = False) #降序排列
counts_['percentage'] = (counts_['num']/counts_['num'].sum())*100
print(counts_)
type num percentage
101 411665 49.156965
102 17357 2.072601
103 1715 0.204788
106 3957 0.472506
107 182900 21.840110
199 201426 24.052302
301 18430 2.200728
代码三:知识类型内部统计
统计107类别的情况
def count107(i): #自定义统计函数
j = i[['fullURL']][i['fullURLId'].str.contains('107')].copy() #找出类别包含107的网址
j['type'] = None #添加空列
j['type'][j['fullURL'].str.contains('info/.+?/')] = u'知识首页' #info以/结尾
j['type'][j['fullURL'].str.contains('info/.+?/.+?')] = u'知识列表页'
j['type'][j['fullURL'].str.contains('/\d+?_*\d+?.html')] = u'知识内容页'
return j['type'].value_counts()
engine = create_engine('mysql+pymysql://root:102011@localhost/test?charset=utf8')
sql = pd.read_sql('all_gzdata', engine, chunksize = 10000)
counts2 = [count107(i) for i in sql] #逐块统计
counts2 = pd.concat(counts2).groupby(level=0).sum() #合并统计结果
print(counts2)
计算各个部分的占比
res107 = pd.DataFrame(counts2)
res107.reset_index(inplace=True)
res107.index.name= u'107类型'
res107.rename(columns={'type':'num'},inplace=True)
res107[u'比例'] = (res107['num']/res107['num'].sum())*100
res107.reset_index(inplace = True)
print(res107)
知识内容页 164243
知识列表页 9656
知识首页 9001
Name: type, dtype: int64
107类型 num 比例
0 知识内容页 164243 89.799344
1 知识列表页 9656 5.279388
2 知识首页 9001 4.921268
代码四:统计带“?”的数据
def countquestion(i): # 自定义统计函数
j = i[['fullURLId']][i['fullURL'].str.contains('?')].copy() # 找出类别包含107的网址
return j
engine = create_engine('mysql+pymysql://root:[email protected]:3306/test?charset=utf8')
sql = pd.read_sql('all_gzdata', engine, chunksize = 10000)
counts3 = [countquestion(i)['fullURLId'].value_counts() for i in sql]
counts3 = pd.concat(counts3).groupby(level=0).sum()
print(counts3)
求各个类型的占比并保存数据
df1 = pd.DataFrame(counts3)
df1['perc'] = df1['fullURLId']/df1['fullURLId'].sum()*100
df1.sort_values(by='fullURLId',ascending=False,inplace=True)
print(df1.round(4))
101003 47
102002 25
107001 346
1999001 64718
301001 356
Name: fullURLId, dtype: int64
fullURLId perc
1999001 64718 98.8182
301001 356 0.5436
107001 346 0.5283
101003 47 0.0718
102002 25 0.0382
复制代码
代码五:统计199类型中的具体类型占比
def page199(i): # 自定义统计函数
j = i[['fullURL', 'pageTitle']][(i['fullURLId'].str.contains('199')) &
(i['fullURL'].str.contains('?'))]
j['pageTitle'].fillna('空', inplace=True)
j['type'] = '其他' # 添加空列
j['type'][j['pageTitle'].str.contains('法律快车-律师助手')] = '法律快车-律师助手'
j['type'][j['pageTitle'].str.contains('咨询发布成功')] = '咨询发布成功'
j['type'][j['pageTitle'].str.contains('免费发布法律咨询')] = '免费发布法律咨询'
j['type'][j['pageTitle'].str.contains('法律快搜')] = '快搜'
j['type'][j['pageTitle'].str.contains('法律快车法律经验')] = '法律快车法律经验'
j['type'][j['pageTitle'].str.contains('法律快车法律咨询')] = '法律快车法律咨询'
j['type'][(j['pageTitle'].str.contains('_法律快车')) |
(j['pageTitle'].str.contains('-法律快车'))] = '法律快车'
j['type'][j['pageTitle'].str.contains('空')] = '空'
return j
注意:获取一次sql对象就需要重新访问一下数据库
engine = create_engine('mysql+pymysql://root:[email protected]:3306/test?charset=utf8')
sql = pd.read_sql('all_gzdata', engine, chunksize=10000) # 分块读取数据库信息
sql = pd.read_sql_query('select * from all_gzdata limit 10000', con=engine)
counts4 = [page199(i) for i in sql] # 逐块统计
counts4 = pd.concat(counts4)
d1 = counts4['type'].value_counts()
print(d1)
d2 = counts4[counts4['type'] == '其他']
print(d2)
求各个部分的占比并保存数据
df1_ = pd.DataFrame(d1)
df1_['perc'] = df1_['type'] / df1_['type'].sum() * 100
df1_.sort_values(by='type', ascending=False, inplace=True)
print(df1_)
法律快车-律师助手 49894
法律快车法律咨询 6421
咨询发布成功 5220
快搜 1943
法律快车 818
其他 359
法律快车法律经验 59
空 4
Name: type, dtype: int64
fullURL
2631 http://www.lawtime.cn/spelawyer/index.php?py=g...
2632 http://www.lawtime.cn/spelawyer/index.php?py=g...
1677 http://m.baidu.com/from=844b/bd_page_type=1/ss...
4303 http://m.baidu.com/from=0/bd_page_type=1/ssid=...
3673 http://www.lawtime.cn/lawyer/lll25879862593080...
... ...
4829 http://www.lawtime.cn/spelawyer/index.php?m=se...
4837 http://www.lawtime.cn/spelawyer/index.php?m=se...
4842 http://www.lawtime.cn/spelawyer/index.php?m=se...
8302 http://www.lawtime.cn/spelawyer/index.php?m=se...
5034 http://www.baidu.com/link?url=O7iBD2KmoJdkHWTZ...
pageTitle type
2631 个旧律师成功案例 - 法律快车提供个旧知名律师、优秀律师、专业律师的咨询和推荐 其他
2632 个旧律师成功案例 - 法律快车提供个旧知名律师、优秀律师、专业律师的咨询和推荐 其他
1677 婚姻法论文 - 法律快车法律论文 其他
4303 什么是机动车?什么是非机动车? - 法律快车交通事故 其他
3673 404错误提示页面 - 法律快车 其他
... ... ...
4829 律师搜索,律师查找 - 法律快车提供全国知名律师、优秀律师、专业律师的咨询和推荐 其他
4837 律师搜索,律师查找 - 法律快车提供全国知名律师、优秀律师、专业律师的咨询和推荐 其他
4842 律师搜索,律师查找 - 法律快车提供全国知名律师、优秀律师、专业律师的咨询和推荐 其他
8302 律师搜索,律师查找 - 法律快车提供全国知名律师、优秀律师、专业律师的咨询和推荐 其他
5034 离婚协议书范本(2015年版) - 法律快车婚姻法 其他
[359 rows x 3 columns]
type 1999001 总数 perc
法律快车-律师助手 49894 77.094471
法律快车法律咨询 6421 9.921506
咨询发布成功 5220 8.065762
快搜 1943 3.002256
法律快车 818 1.263945
其他 359 0.554714
法律快车法律经验 59 0.091165
空 4 0.006181
代码六:统计无目的浏览用户中各个类型占比
6无目的用户
def xiaguang(i): # 自定义统计函数
j = i.loc[(i['fullURL'].str.contains('.html')) == False,
['fullURL', 'fullURLId', 'pageTitle']]
return j
注意获取一次sql对象就需要重新访问一下数据库
engine = create_engine('mysql+pymysql://root:[email protected]:3306/test?charset=utf8')
sql = pd.read_sql('all_gzdata', engine, chunksize=10000) # 分块读取数据库信息
counts5 = [xiaguang(i) for i in sql]
counts5 = pd.concat(counts5)
xg1 = counts5['fullURLId'].value_counts()
print(xg1)
求各个部分的占比
xg_ = pd.DataFrame(xg1)
xg_.reset_index(inplace=True)
xg_.columns = ['index', 'num']
xg_['perc'] = xg_['num'] / xg_['num'].sum() * 100
xg_.sort_values(by='num', ascending=False, inplace=True)
xg_['type'] = xg_['index'].str.extract('(\d{3})') # 提取前三个数字作为类别id
xgs_ = xg_[['type', 'num']].groupby('type').sum() # 按类别合并
xgs_.sort_values(by='num', ascending=False, inplace=True) # 降序排列
xgs_['percentage'] = xgs_['num'] / xgs_['num'].sum() * 100
print(xgs_.round(4))
1999001 117124
107001 17843
102002 12021
101001 5603
106001 3957
102001 2129
102003 1235
301001 1018
101009 854
102007 538
102008 404
101008 378
102004 361
102005 271
102009 214
102006 184
101004 125
101006 107
101005 63
Name: fullURLId, dtype: int64
type num percentage
199 117124 71.2307
107 17843 10.8515
102 17357 10.5559
101 7130 4.3362
106 3957 2.4065
301 1018 0.6191
代码七:统计用户浏览网页次数情况
分析网页点击次数
统计点击次数
engine = create_engine('mysql+pymysql://root:[email protected]:3306/test?charset=utf8')
sql = pd.read_sql('all_gzdata', engine, chunksize = 10000)# 分块读取数据库信息
counts1 = [i['realIP'].value_counts() for i in sql] # 分块统计各个IP的出现次数
counts1 = pd.concat(counts1).groupby(level=0).sum() # 合并统计结果,level=0表示按照index分组
print(counts1)
counts1_ = pd.DataFrame(counts1)
counts1_
counts1['realIP'] = counts1.index.tolist()
counts1_[1]=1 # 添加1列全为1
hit_count = counts1_.groupby('realIP').sum() # 统计各个“不同点击次数”分别出现的次数
也可以使用counts1_['realIP'].value_counts()功能
hit_count.columns=['用户数']
hit_count.index.name = '点击次数'
统计1~7次、7次以上的用户人数
hit_count.sort_index(inplace = True)
hit_count_7 = hit_count.iloc[:7,:]
time = hit_count.iloc[7:,0].sum() # 统计点击次数7次以上的用户数
hit_count_7 = hit_count_7.append([{'用户数':time}], ignore_index=True)
hit_count_7.index = ['1','2','3','4','5','6','7','7次以上']
hit_count_7['用户比例'] = hit_count_7['用户数'] / hit_count_7['用户数'].sum()
print(hit_count_7)
82033 2
95502 1
103182 1
116010 2
136206 1
..
4294809358 2
4294811150 1
4294852154 3
4294865422 2
4294917690 1
Name: realIP, Length: 230149, dtype: int64
用户数 用户比例
1 132119 0.574059
2 44175 0.191941
3 17573 0.076355
4 10156 0.044128
5 5952 0.025862
6 4132 0.017954
7 2632 0.011436
7次以上 13410 0.058267
代码八:分析浏览次数为一次的用户的行为
分析浏览一次的用户行为
engine = create_engine('mysql+pymysql://root:[email protected]:3306/test?charset=utf8')
all_gzdata = pd.read_sql_table('all_gzdata', con = engine) # 读取all_gzdata数据
对realIP进行统计
提取浏览1次网页的数据
real_count = pd.DataFrame(all_gzdata.groupby("realIP")["realIP"].count())
real_count.columns = ["count"]
real_count["realIP"] = real_count.index.tolist()
user_one = real_count[(real_count["count"] == 1)] # 提取只登录一次的用户
通过realIP与原始数据合并
real_one = pd.merge(user_one, all_gzdata,right_on='realIP',left_index=True,how='left')
统计浏览一次的网页类型
URL_count = pd.DataFrame(real_one.groupby("fullURLId")["fullURLId"].count())
URL_count.columns = ["count"]
URL_count.sort_values(by='count', ascending=False, inplace=True) # 降序排列
统计排名前4和其他的网页类型
URL_count_4 = URL_count.iloc[:4,:]
time = hit_count.iloc[4:,0].sum() # 统计其他的
URLindex = URL_count_4.index.values
URL_count_4 = URL_count_4.append([{'count':time}], ignore_index=True)
URL_count_4.index = [URLindex[0], URLindex[1], URLindex[2], URLindex[3],
'其他']
URL_count_4['比例'] = URL_count_4['count'] / URL_count_4['count'].sum()
print(URL_count_4)
count 比例
101003 102560 0.649011
107001 19443 0.123037
1999001 9381 0.059364
301001 515 0.003259
其他 26126 0.165328
代码九:统计单用户浏览次数为一的网页
在浏览1次的前提下, 得到的网页被浏览的总次数
fullURL_count = pd.DataFrame(real_one.groupby("fullURL")["fullURL"].count())
fullURL_count.columns = ["count"]
fullURL_count["fullURL"] = fullURL_count.index.tolist()
fullURL_count.sort_values(by='count', ascending=False, inplace=True) # 降序排列
print(fullURL_count.head(10))
count \
fullURL
http://www.lawtime.cn/info/shuifa/slb/201211197... 1013
http://www.lawtime.cn/info/hunyin/lhlawlhxy/201... 501
http://www.lawtime.cn/ask/question_925675.html 423
http://www.lawtime.cn/info/shuifa/slb/201211197... 367
http://www.lawtime.cn/ask/exp/13655.html 301
http://www.lawtime.cn/ask/exp/8495.html 241
http://www.lawtime.cn/ask/exp/13445.html 199
http://www.lawtime.cn/guangzhou 177
http://www.lawtime.cn/ask/exp/17357.html 171
http://www.lawtime.cn/citylist.php 117
fullURL
fullURL
http://www.lawtime.cn/info/shuifa/slb/201211197... http://www.lawtime.cn/info/shuifa/slb/20121119...
http://www.lawtime.cn/info/hunyin/lhlawlhxy/201... http://www.lawtime.cn/info/hunyin/lhlawlhxy/20...
http://www.lawtime.cn/ask/question_925675.html http://www.lawtime.cn/ask/question_925675.html
http://www.lawtime.cn/info/shuifa/slb/201211197... http://www.lawtime.cn/info/shuifa/slb/20121119...
http://www.lawtime.cn/ask/exp/13655.html http://www.lawtime.cn/ask/exp/13655.html
http://www.lawtime.cn/ask/exp/8495.html http://www.lawtime.cn/ask/exp/8495.html
http://www.lawtime.cn/ask/exp/13445.html http://www.lawtime.cn/ask/exp/13445.html
http://www.lawtime.cn/guangzhou http://www.lawtime.cn/guangzhou
http://www.lawtime.cn/ask/exp/17357.html http://www.lawtime.cn/ask/exp/17357.html
http://www.lawtime.cn/citylist.php http://www.lawtime.cn/citylist.php
第二部分——数据预处理
代码十:删除不符合规则的网页
import os
import re
import pandas as pd
import pymysql as pm
from random import sample
修改工作路径到指定文件夹
os.chdir("D:/python123")
读取数据
con = pm.connect(host='localhost',user='root',password='102011',database='test',charset='utf8')
data = pd.read_sql('select * from all_gzdata',con=con)
con.close() #关闭连接
取出107类型数据
index107 = [re.search('107',str(i))!=None for i in data.loc[:,'fullURLId']]
data_107 = data.loc[index107,:]
在107类型中筛选出婚姻类数据
index = [re.search('hunyin',str(i))!=None for i in data_107.loc[:,'fullURL']]
data_hunyin = data_107.loc[index,:]
提取所需字段(realIP、fullURL)
info = data_hunyin.loc[:,['realIP','fullURL']]
去除网址中“?”及其后面内容
da = [re.sub('?.*','',str(i)) for i in info.loc[:,'fullURL']]
info.loc[:,'fullURL'] = da # 将info中‘fullURL’那列换成da
去除无html网址
index = [re.search('.html',str(i))!=None for i in info.loc[:,'fullURL']]
index.count(True) # True 或者 1 , False 或者 0
info1 = info.loc[index,:]
print(info1.head())
realIP fullURL
0 2683657840 http://www.lawtime.cn/info/hunyin/hunyinfagui/...
4 2683657840 http://www.lawtime.cn/info/hunyin/hunyinfagui/...
9 1275347569 http://www.lawtime.cn/info/hunyin/lhlawlhxy/20...
62 1531496412 http://www.lawtime.cn/info/hunyin/hunyinfagui/...
86 838215995 http://www.lawtime.cn/info/hunyin/lhlawlhxy/20...
代码十一:还原翻页地址
找出翻页和非翻页网址
index = [re.search('/\d+_\d+.html',i)!=None for i in info1.loc[:,'fullURL']]
index1 = [i==False for i in index]
info1_1 = info1.loc[index,:] # 带翻页网址
info1_2 = info1.loc[index1,:] # 无翻页网址
将翻页网址还原
da = [re.sub('_\d+.html','.html',str(i)) for i in info1_1.loc[:,'fullURL']]
info1_1.loc[:,'fullURL'] = da
翻页与非翻页网址合并
frames = [info1_1,info1_2]
info2 = pd.concat(frames)
或者
info2 = pd.concat([info1_1,info1_2],axis = 0) # 默认为0,即行合并
去重(realIP和fullURL两列相同)
info3 = info2.drop_duplicates()
将IP转换成字符型数据
info3.iloc[:,0] = [str(index) for index in info3.iloc[:,0]]
info3.iloc[:,1] = [str(index) for index in info3.iloc[:,1]]
len(info3)
代码十二:筛去浏览次数不满两次的用户
筛选满足一定浏览次数的IP
IP_count = info3['realIP'].value_counts()
找出IP集合
IP = list(IP_count.index)
count = list(IP_count.values)
统计每个IP的浏览次数,并存放进IP_count数据框中,第一列为IP,第二列为浏览次数
IP_count = pd.DataFrame({'IP':IP,'count':count})
3.3筛选出浏览网址在n次以上的IP集合
n = 2
index = IP_count.loc[:,'count']>n
IP_index = IP_count.loc[index,'IP']
print(IP_index.head())
0 2609113527
1 3812410744
2 225896631
3 242673847
4 1190924814
Name: IP, dtype: object
代码十三:划分数据集
划分IP集合为训练集和测试集
index_tr = sample(range(0,len(IP_index)),int(len(IP_index)*0.8)) # 或者np.random.sample
index_te = [i for i in range(0,len(IP_index)) if i not in index_tr]
IP_tr = IP_index[index_tr]
IP_te = IP_index[index_te]
将对应数据集划分为训练集和测试集
index_tr = [i in list(IP_tr) for i in info3.loc[:,'realIP']]
index_te = [i in list(IP_te) for i in info3.loc[:,'realIP']]
data_tr = info3.loc[index_tr,:]
data_te = info3.loc[index_te,:]
print(len(data_tr))
IP_tr = data_tr.iloc[:,0] # 训练集IP
url_tr = data_tr.iloc[:,1] # 训练集网址
IP_tr = list(set(IP_tr)) # 去重处理
url_tr = list(set(url_tr)) # 去重处理
len(url_tr)
代码十四:构建模型
利用训练集数据构建模型
UI_matrix_tr = pd.DataFrame(0,index=IP_tr,columns=url_tr)
求用户-物品矩阵
for i in data_tr.index:
UI_matrix_tr.loc[data_tr.loc[i,'realIP'],data_tr.loc[i,'fullURL']] = 1
sum(UI_matrix_tr.sum(axis=1))
求物品相似度矩阵(因计算量较大,需要耗费的时间较久)
Item_matrix_tr = pd.DataFrame(0,index=url_tr,columns=url_tr)
for i in Item_matrix_tr.index:
for j in Item_matrix_tr.index:
a = sum(UI_matrix_tr.loc[:,[i,j]].sum(axis=1)==2)
b = sum(UI_matrix_tr.loc[:,[i,j]].sum(axis=1)!=0)
Item_matrix_tr.loc[i,j] = a/b
将物品相似度矩阵对角线处理为零
for i in Item_matrix_tr.index:
Item_matrix_tr.loc[i,i]=0
利用测试集数据对模型评价
IP_te = data_te.iloc[:,0]
url_te = data_te.iloc[:,1]
IP_te = list(set(IP_te))
url_te = list(set(url_te))
测试集数据用户物品矩阵
UI_matrix_te = pd.DataFrame(0,index=IP_te,columns=url_te)
for i in data_te.index:
UI_matrix_te.loc[data_te.loc[i,'realIP'],data_te.loc[i,'fullURL']] = 1
对测试集IP进行推荐
Res = pd.DataFrame('NaN',index=data_te.index,columns=['IP','已浏览网址','推荐网址','T/F'])
Res.loc[:,'IP']=list(data_te.iloc[:,0])
Res.loc[:,'已浏览网址']=list(data_te.iloc[:,1])
开始推荐
for i in Res.index:
if Res.loc[i,'已浏览网址'] in list(Item_matrix_tr.index):
Res.loc[i,'推荐网址'] = Item_matrix_tr.loc[Res.loc[i,'已浏览网址'],:].idxmax()
if Res.loc[i,'推荐网址'] in url_te:
Res.loc[i,'T/F']=UI_matrix_te.loc[Res.loc[i,'IP'],Res.loc[i,'推荐网址']]==1
else:
Res.loc[i,'T/F'] = False
保存推荐结果
Res.to_csv('D:/python123/Res.csv',index=False,encoding='utf8')
代码十五:计算推荐结果的准确率、召回率和F1指标
读取保存的推荐结果
Res = pd.read_csv('D:/python123/Res.csv',keep_default_na=False, encoding='utf8')
计算推荐准确率
Pre = round(sum(Res.loc[:,'T/F']'True') / (len(Res.index)-sum(Res.loc[:,'T/F']'NaN')), 3)
print('推荐准确率:',Pre)
计算推荐召回率
Rec = round(sum(Res.loc[:,'T/F']'True') / (sum(Res.loc[:,'T/F']'True')+sum(Res.loc[:,'T/F']=='NaN')), 3)
print('推荐召回率:',Rec)
计算F1指标
F1 = round(2PreRec/(Pre+Rec),3)
print('推荐F1指标:',F1)