机器学习sklearn（73）：算法实例（三十）分类（十七）SVM（八）sklearn.svm.SVC（七） SVC真实数据案例：预测明天是否会下雨
0 简介
Kaggle下载链接走这里：https://www.kaggle.com/jsphyg/weather-dataset-rattle-package
1 导库导数据，探索特征
导入需要的库
import pandas as pd import numpy as np from sklearn.model_selection import train_test_split
导入数据，探索数据
weather = pd.read_csv(r"C:\work\learnbetter\micro-class\week 8SVM (2)\data\weatherAUS5000.csv",index_col=0) weather.head()
来查看一下各个特征都代表了什么：
#将特征矩阵和标签Y分开 X = weather.iloc[:,:-1] Y = weather.iloc[:,-1] X.shape #探索数据类型 X.info() #探索缺失值 X.isnull().mean() #探索标签的分类 np.unique(Y)
2 分集，优先探索标签
分训练集和测试集，并做描述性统计
#分训练集和测试集 Xtrain, Xtest, Ytrain, Ytest = train_test_split(X,Y,test_size=0.3,random_state=420) #恢复索引 for i in [Xtrain, Xtest, Ytrain, Ytest]: i.index = range(i.shape[0])
#是否有样本不平衡问题？ Ytrain.value_counts() Ytest.value_counts() #将标签编码 from sklearn.preprocessing import LabelEncoder encorder = LabelEncoder().fit(Ytrain) Ytrain = pd.DataFrame(encorder.transform(Ytrain)) Ytest = pd.DataFrame(encorder.transform(Ytest))
3 探索特征，开始处理特征矩阵
3.1 描述性统计与异常值
#描述性统计 Xtrain.describe([0.01,0.05,0.1,0.25,0.5,0.75,0.9,0.99]).T Xtest.describe([0.01,0.05,0.1,0.25,0.5,0.75,0.9,0.99]).T """ 对于去kaggle上下载了数据的小伙伴们，以及坚持要使用完整版数据的（15W行）小伙伴们 如果你发现了异常值，首先你要观察，这个异常值出现的频率 如果异常值只出现了一次，多半是输入错误，直接把异常值删除 如果异常值出现了多次，去跟业务人员沟通，可能这是某种特殊表示，如果是人为造成的错误，异常值留着是没有用 的，只要数据量不是太大，都可以删除 如果异常值占到你总数据量的10%以上了，不能轻易删除。可以考虑把异常值替换成非异常但是非干扰的项，比如说用0 来进行替换，或者把异常当缺失值，用均值或者众数来进行替换 """ #先查看原始的数据结构 Xtrain.shape Xtest.shape #观察异常值是大量存在，还是少数存在 Xtrain.loc[Xtrain.loc[:,"Cloud9am"] == 9,"Cloud9am"] Xtest.loc[Xtest.loc[:,"Cloud9am"] == 9,"Cloud9am"] Xtest.loc[Xtest.loc[:,"Cloud3pm"] == 9,"Cloud3pm"] #少数存在，于是采取删除的策略 #注意如果删除特征矩阵，则必须连对应的标签一起删除，特征矩阵的行和标签的行必须要一一对应 Xtrain = Xtrain.drop(index = 71737) Ytrain = Ytrain.drop(index = 71737) #删除完毕之后，观察原始的数据结构，确认删除正确 Xtrain.shape Xtest = Xtest.drop(index = [19646,29632]) Ytest = Ytest.drop(index = [19646,29632]) Xtest.shape #进行任何行删除之后，千万记得要恢复索引 for i in [Xtrain, Xtest, Ytrain, Ytest]: i.index = range(i.shape[0]) Xtrain.head() Xtest.head()
3.2 处理困难特征：日期
我们采集数据的日期是否和我们的天气有关系呢？我们可以探索一下我们的采集日期有什么样的性质：
Xtrainc = Xtrain.copy() Xtrainc.sort_values(by="Location") Xtrain.iloc[:,0].value_counts() #首先，日期不是独一无二的，日期有重复 #其次，在我们分训练集和测试集之后，日期也不是连续的，而是分散的 #某一年的某一天倾向于会下雨？或者倾向于不会下雨吗？ #不是日期影响了下雨与否，反而更多的是这一天的日照时间，湿度，温度等等这些因素影响了是否会下雨 #光看日期，其实感觉它对我们的判断并无直接影响 #如果我们把它当作连续型变量处理，那算法会人为它是一系列1~3000左右的数字，不会意识到这是日期 Xtrain.iloc[:,0].value_counts().count() #如果我们把它当作分类型变量处理，类别太多，有2141类，如果换成数值型，会被直接当成连续型变量，如果做成哑变量，我们特征的维度会爆炸
Xtrain = Xtrain.drop(["Date"],axis=1) Xtest = Xtest.drop(["Date"],axis=1)
Xtrain["Rainfall"].head(20) Xtrain.loc[Xtrain["Rainfall"] >= 1,"RainToday"] = "Yes" Xtrain.loc[Xtrain["Rainfall"] < 1,"RainToday"] = "No" Xtrain.loc[Xtrain["Rainfall"] == np.nan,"RainToday"] = np.nan Xtest.loc[Xtest["Rainfall"] >= 1,"RainToday"] = "Yes" Xtest.loc[Xtest["Rainfall"] < 1,"RainToday"] = "No" Xtest.loc[Xtest["Rainfall"] == np.nan,"RainToday"] = np.nan Xtrain.head() Xtest.head()
int(Xtrain.loc[0,"Date"].split("-")[1]) #提取出月份 Xtrain["Date"] = Xtrain["Date"].apply(lambda x:int(x.split("-")[1])) #替换完毕后，我们需要修改列的名称 #rename是比较少有的，可以用来修改单个列名的函数 #我们通常都直接使用 df.columns = 某个列表 这样的形式来一次修改所有的列名 #但rename允许我们只修改某个单独的列 Xtrain = Xtrain.rename(columns={"Date":"Month"}) Xtrain.head() Xtest["Date"] = Xtest["Date"].apply(lambda x:int(x.split("-")[1])) Xtest = Xtest.rename(columns={"Date":"Month"}) Xtest.head()
3.3 处理困难特征：地点
import time from selenium import webdriver #导入需要的模块，其中爬虫使用的是selenium import pandas as pd import numpy as np df = pd.DataFrame(index=range(len(cityname))) #创建新dataframe用于存储爬取的数据 driver = webdriver.Chrome() #调用谷歌浏览器 time0 = time.time() #计时开始 #循环开始 for num, city in enumerate(cityname): #在城市名称中进行遍历 driver.get('https://www.google.co.uk/webhp? hl=en&sa=X&ved=0ahUKEwimtcX24cTfAhUJE7wKHVkWB5AQPAgH') #首先打开谷歌主页 time.sleep(0.3) #停留0.3秒让我们知道发生了什么 search_box = driver.find_element_by_name('q') #锁定谷歌的搜索输入框 search_box.send_keys('%s Australia Latitude and longitude' % (city)) #在输入框中输入 “城市” 澳大利亚 经纬度 search_box.submit() #enter，确认开始搜索 result = driver.find_element_by_xpath('//div[@class="Z0LcW"]').text #？爬取需要的经纬 度，就是这里，怎么获取的呢？ resultsplit = result.split(" ") #将爬取的结果用split进行分割 df.loc[num,"City"] = city #向提前创建好的df中输入爬取的数据，第一列是城市名 df.loc[num,"Latitude"] = resultsplit[0] #第二列是纬度 df.loc[num,"Longitude"] = resultsplit[2] #第三列是经度 df.loc[num,"Latitudedir"] = resultsplit[1] #第四列是纬度的方向 df.loc[num,"Longitudedir"] = resultsplit[3] #第五列是经度的方向 print("%i webcrawler successful for city %s" % (num,city)) #每次爬虫成功之后，就打印“城 市”成功了 time.sleep(1) #全部爬取完毕后，停留1秒钟 driver.quit() #关闭浏览器 print(time.time() - time0) #打印所需的时间
cityll = pd.read_csv(r"C:\work\learnbetter\micro-class\week 8 SVM (2)\cityll.csv",index_col=0) city_climate = pd.read_csv(r"C:\work\learnbetter\micro-class\week 8 SVM (2)\Cityclimate.csv") cityll.head() city_climate.head()
#去掉度数符号 cityll["Latitudenum"] = cityll["Latitude"].apply(lambda x:float(x[:-1])) cityll["Longitudenum"] = cityll["Longitude"].apply(lambda x:float(x[:-1])) #观察一下所有的经纬度方向都是一致的，全部是南纬，东经，因为澳大利亚在南半球，东半球 #所以经纬度的方向我们可以舍弃了 citylld = cityll.iloc[:,[0,5,6]] #将city_climate中的气候添加到我们的citylld中 citylld["climate"] = city_climate.iloc[:,-1] citylld.head()
#训练集中所有的地点 cityname = Xtrain.iloc[:,1].value_counts().index.tolist() cityname import time from selenium import webdriver #导入需要的模块，其中爬虫使用的是selenium import pandas as pd import numpy as np df = pd.DataFrame(index=range(len(cityname))) #创建新dataframe用于存储爬取的数据 driver = webdriver.Chrome() #调用谷歌浏览器 time0 = time.time() #计时开始 #循环开始 for num, city in enumerate(cityname): #在城市名称中进行遍历 driver.get('https://www.google.co.uk/webhp? hl=en&sa=X&ved=0ahUKEwimtcX24cTfAhUJE7wKHVkWB5AQPAgH') #首先打开谷歌主页 time.sleep(0.3) #停留0.3秒让我们知道发生了什么 search_box = driver.find_element_by_name('q') #锁定谷歌的搜索输入框 search_box.send_keys('%s Australia Latitude and longitude' % (city)) #在输入框中输入 “城市” 澳大利亚 经纬度 search_box.submit() #enter，确认开始搜索 result = driver.find_element_by_xpath('//div[@class="Z0LcW"]').text #？爬取需要的经纬 度，就是这里，怎么获取的呢？ resultsplit = result.split(" ") #将爬取的结果用split进行分割 df.loc[num,"City"] = city #向提前创建好的df中输入爬取的数据，第一列是城市名 df.loc[num,"Latitude"] = resultsplit[0] #第二列是经度 df.loc[num,"Longitude"] = resultsplit[2] #第三列是纬度 df.loc[num,"Latitudedir"] = resultsplit[1] #第四列是经度的方向 df.loc[num,"Longitudedir"] = resultsplit[3] #第五列是纬度的方向 print("%i webcrawler successful for city %s" % (num,city)) #每次爬虫成功之后，就打印“城 市”成功了 time.sleep(1) #全部爬取完毕后，停留1秒钟 driver.quit() #关闭浏览器 print(time.time() - time0) #打印所需的时间 df.to_csv(r"C:\work\learnbetter\micro-class\week 8 SVM (2)\samplecity.csv")
来查看一下我们爬取出的内容是什么样子：
samplecity = pd.read_csv(r"C:\work\learnbetter\micro-class\week 8 SVM (2)\samplecity.csv",index_col=0) #我们对samplecity也执行同样的处理：去掉经纬度中度数的符号，并且舍弃我们的经纬度的方向 samplecity["Latitudenum"] = samplecity["Latitude"].apply(lambda x:float(x[:-1])) samplecity["Longitudenum"] = samplecity["Longitude"].apply(lambda x:float(x[:-1])) samplecityd = samplecity.iloc[:,[0,5,6]] samplecityd.head()
#首先使用radians将角度转换成弧度 from math import radians, sin, cos, acos citylld.loc[:,"slat"] = citylld.iloc[:,1].apply(lambda x : radians(x)) citylld.loc[:,"slon"] = citylld.iloc[:,2].apply(lambda x : radians(x)) samplecityd.loc[:,"elat"] = samplecityd.iloc[:,1].apply(lambda x : radians(x)) samplecityd.loc[:,"elon"] = samplecityd.iloc[:,2].apply(lambda x : radians(x)) import sys for i in range(samplecityd.shape[0]): slat = citylld.loc[:,"slat"] slon = citylld.loc[:,"slon"] elat = samplecityd.loc[i,"elat"] elon = samplecityd.loc[i,"elon"] dist = 6371.01 * np.arccos(np.sin(slat)*np.sin(elat) + np.cos(slat)*np.cos(elat)*np.cos(slon.values - elon)) city_index = np.argsort(dist)[0] #每次计算后，取距离最近的城市，然后将最近的城市和城市对应的气候都匹配到samplecityd中 samplecityd.loc[i,"closest_city"] = citylld.loc[city_index,"City"] samplecityd.loc[i,"climate"] = citylld.loc[city_index,"climate"] #查看最后的结果，需要检查城市匹配是否基本正确 samplecityd.head() #查看气候的分布 samplecityd["climate"].value_counts() #确认无误后，取出样本城市所对应的气候，并保存 locafinal = samplecityd.iloc[:,[0,-1]] locafinal.head() locafinal.columns = ["Location","Climate"] #在这里设定locafinal的索引为地点，是为了之后进行map的匹配 locafinal = locafinal.set_index(keys="Location") locafinal.to_csv(r"C:\work\learnbetter\micro-class\week 8 SVM (2)\samplelocation.csv") locafinal.head()
#是否还记得训练集长什么样呢？ Xtrain.head() #将location中的内容替换，并且确保匹配进入的气候字符串中不含有逗号，气候两边不含有空格 #我们使用re这个模块来消除逗号 #re.sub(希望替换的值，希望被替换成的值，要操作的字符串) #x.strip()是去掉空格的函数 import re Xtrain["Location"] = Xtrain["Location"].map(locafinal.iloc[:,0]).apply(lambda x:re.sub(",","",x.strip())) Xtest["Location"] = Xtest["Location"].map(locafinal.iloc[:,0]).apply(lambda x:re.sub(",","",x.strip())) #修改特征内容之后，我们使用新列名“Climate”来替换之前的列名“Location” #注意这个命令一旦执行之后，就再没有列"Location"了，使用索引时要特别注意 Xtrain = Xtrain.rename(columns={"Location":"Climate"}) Xtest = Xtest.rename(columns={"Location":"Climate"}) Xtrain.head() Xtest.head()
3.4 处理分类型变量：缺失值
#查看缺失值的缺失情况 Xtrain.isnull().mean() #首先找出，分类型特征都有哪些 cate = Xtrain.columns[Xtrain.dtypes == "object"].tolist() #除了特征类型为"object"的特征们，还有虽然用数字表示，但是本质为分类型特征的云层遮蔽程度 cloud = ["Cloud9am","Cloud3pm"] cate = cate + cloud cate #对于分类型特征，我们使用众数来进行填补 from sklearn.impute import SimpleImputer si = SimpleImputer(missing_values=np.nan,strategy="most_frequent") #注意，我们使用训练集数据来训练我们的填补器，本质是在生成训练集中的众数 si.fit(Xtrain.loc[:,cate]) #然后我们用训练集中的众数来同时填补训练集和测试集 Xtrain.loc[:,cate] = si.transform(Xtrain.loc[:,cate]) Xtest.loc[:,cate] = si.transform(Xtest.loc[:,cate]) Xtrain.head() Xtest.head() #查看分类型特征是否依然存在缺失值 Xtrain.loc[:,cate].isnull().mean() Xtest.loc[:,cate].isnull().mean()
3.5 处理分类型变量：将分类型变量编码
#将所有的分类型变量编码为数字，一个类别是一个数字 from sklearn.preprocessing import OrdinalEncoder oe = OrdinalEncoder() #利用训练集进行fit oe = oe.fit(Xtrain.loc[:,cate]) #用训练集的编码结果来编码训练和测试特征矩阵 #在这里如果测试特征矩阵报错，就说明测试集中出现了训练集中从未见过的类别 Xtrain.loc[:,cate] = oe.transform(Xtrain.loc[:,cate]) Xtest.loc[:,cate] = oe.transform(Xtest.loc[:,cate]) Xtrain.loc[:,cate].head() Xtest.loc[:,cate].head()
3.6 处理连续型变量：填补缺失值
col = Xtrain.columns.tolist() for i in cate: col.remove(i) col #实例化模型，填补策略为"mean"表示均值 impmean = SimpleImputer(missing_values=np.nan,strategy = "mean") #用训练集来fit模型 impmean = impmean.fit(Xtrain.loc[:,col]) #分别在训练集和测试集上进行均值填补 Xtrain.loc[:,col] = impmean.transform(Xtrain.loc[:,col]) Xtest.loc[:,col] = impmean.transform(Xtest.loc[:,col]) Xtrain.head() Xtest.head()
3.7 处理连续型变量：无量纲化
col.remove("Month") col from sklearn.preprocessing import StandardScaler ss = StandardScaler() ss = ss.fit(Xtrain.loc[:,col]) Xtrain.loc[:,col] = ss.transform(Xtrain.loc[:,col]) Xtest.loc[:,col] = ss.transform(Xtest.loc[:,col]) Xtrain.head() Xtest.head() Ytrain.head() Ytest.head()
4 建模与模型评估
from time import time import datetime from sklearn.svm import SVC from sklearn.model_selection import cross_val_score from sklearn.metrics import roc_auc_score, recall_score Ytrain = Ytrain.iloc[:,0].ravel() Ytest = Ytest.iloc[:,0].ravel() #建模选择自然是我们的支持向量机SVC，首先用核函数的学习曲线来选择核函数 #我们希望同时观察，精确性，recall以及AUC分数 times = time() #因为SVM是计算量很大的模型，所以我们需要时刻监控我们的模型运行时间 for kernel in ["linear","poly","rbf","sigmoid"]: clf = SVC(kernel = kernel ,gamma="auto" ,degree = 1 ,cache_size = 5000 ).fit(Xtrain, Ytrain) result = clf.predict(Xtest) score = clf.score(Xtest,Ytest) recall = recall_score(Ytest, result) auc = roc_auc_score(Ytest,clf.decision_function(Xtest)) print("%s 's testing accuracy %f, recall is %f', auc is %f" % (kernel,score,recall,auc)) print(datetime.datetime.fromtimestamp(time()-times).strftime("%M:%S:%f"))
5 模型调参
5.1 最求最高Recall
times = time() for kernel in ["linear","poly","rbf","sigmoid"]: clf = SVC(kernel = kernel ,gamma="auto" ,degree = 1 ,cache_size = 5000 ,class_weight = "balanced" ).fit(Xtrain, Ytrain) result = clf.predict(Xtest) score = clf.score(Xtest,Ytest) recall = recall_score(Ytest, result) auc = roc_auc_score(Ytest,clf.decision_function(Xtest)) print("%s 's testing accuracy %f, recall is %f', auc is %f" % (kernel,score,recall,auc)) print(datetime.datetime.fromtimestamp(time()-times).strftime("%M:%S:%f"))
times = time() clf = SVC(kernel = "linear" ,gamma="auto" ,cache_size = 5000 ,class_weight = {1:10} #注意，这里写的其实是，类别1：10，隐藏了类别0：1这个比例 ).fit(Xtrain, Ytrain) result = clf.predict(Xtest) score = clf.score(Xtest,Ytest) recall = recall_score(Ytest, result) auc = roc_auc_score(Ytest,clf.decision_function(Xtest)) print("testing accuracy %f, recall is %f', auc is %f" %(score,recall,auc)) print(datetime.datetime.fromtimestamp(time()-times).strftime("%M:%S:%f"))
5.2 追求最高准确率
valuec = pd.Series(Ytest).value_counts()
valuec
valuec[0]/valuec.sum()
#查看模型的特异度 from sklearn.metrics import confusion_matrix as CM clf = SVC(kernel = "linear" ,gamma="auto" ,cache_size = 5000 ).fit(Xtrain, Ytrain) result = clf.predict(Xtest) cm = CM(Ytest,result,labels=(1,0)) cm specificity = cm[1,1]/cm[1,:].sum() specificity #几乎所有的0都被判断正确了，还有不少1也被判断正确了
irange = np.linspace(0.01,0.05,10) for i in irange: times = time() clf = SVC(kernel = "linear" ,gamma="auto" ,cache_size = 5000 ,class_weight = {1:1+i} ).fit(Xtrain, Ytrain) result = clf.predict(Xtest) score = clf.score(Xtest,Ytest) recall = recall_score(Ytest, result) auc = roc_auc_score(Ytest,clf.decision_function(Xtest)) print("under ratio 1:%f testing accuracy %f, recall is %f', auc is %f" % (1+i,score,recall,auc)) print(datetime.datetime.fromtimestamp(time()-times).strftime("%M:%S:%f"))
irange_ = np.linspace(0.018889,0.027778,10) for i in irange_: times = time() clf = SVC(kernel = "linear" ,gamma="auto" ,cache_size = 5000 ,class_weight = {1:1+i} ).fit(Xtrain, Ytrain) result = clf.predict(Xtest) score = clf.score(Xtest,Ytest) recall = recall_score(Ytest, result) auc = roc_auc_score(Ytest,clf.decision_function(Xtest)) print("under ratio 1:%f testing accuracy %f, recall is %f', auc is %f" % (1+i,score,recall,auc)) print(datetime.datetime.fromtimestamp(time()-times).strftime("%M:%S:%f"))
from sklearn.linear_model import LogisticRegression as LR logclf = LR(solver="liblinear").fit(Xtrain, Ytrain) logclf.score(Xtest,Ytest) C_range = np.linspace(3,5,10) for C in C_range: logclf = LR(solver="liblinear",C=C).fit(Xtrain, Ytrain) print(C,logclf.score(Xtest,Ytest))
5.3 追求平衡
###======【TIME WARNING：10mins】======### import matplotlib.pyplot as plt C_range = np.linspace(0.01,20,20) recallall = [] aucall = [] scoreall = [] for C in C_range: times = time() clf = SVC(kernel = "linear",C=C,cache_size = 5000 ,class_weight = "balanced" ).fit(Xtrain, Ytrain) result = clf.predict(Xtest) score = clf.score(Xtest,Ytest) recall = recall_score(Ytest, result) auc = roc_auc_score(Ytest,clf.decision_function(Xtest)) recallall.append(recall) aucall.append(auc) scoreall.append(score) print("under C %f, testing accuracy is %f,recall is %f', auc is %f" % (C,score,recall,auc)) print(datetime.datetime.fromtimestamp(time()-times).strftime("%M:%S:%f")) print(max(aucall),C_range[aucall.index(max(aucall))]) plt.figure() plt.plot(C_range,recallall,c="red",label="recall") plt.plot(C_range,aucall,c="black",label="auc") plt.plot(C_range,scoreall,c="orange",label="accuracy") plt.legend() plt.show()
times = time() clf = SVC(kernel = "linear",C=3.1663157894736838,cache_size = 5000 ,class_weight = "balanced" ).fit(Xtrain, Ytrain) result = clf.predict(Xtest) score = clf.score(Xtest,Ytest) recall = recall_score(Ytest, result) auc = roc_auc_score(Ytest,clf.decision_function(Xtest)) print("testing accuracy %f,recall is %f', auc is %f" % (score,recall,auc)) print(datetime.datetime.fromtimestamp(time()-times).strftime("%M:%S:%f"))
from sklearn.metrics import roc_curve as ROC import matplotlib.pyplot as plt FPR, Recall, thresholds = ROC(Ytest,clf.decision_function(Xtest),pos_label=1) area = roc_auc_score(Ytest,clf.decision_function(Xtest)) plt.figure() plt.plot(FPR, Recall, color='red', label='ROC curve (area = %0.2f)' % area) plt.plot([0, 1], [0, 1], color='black', linestyle='--') plt.xlim([-0.05, 1.05]) plt.ylim([-0.05, 1.05]) plt.xlabel('False Positive Rate') plt.ylabel('Recall') plt.title('Receiver operating characteristic example') plt.legend(loc="lower right") plt.show()
以此模型作为基础，我们来求解最佳阈值：
maxindex = (Recall - FPR).tolist().index(max(Recall - FPR))
thresholds[maxindex]
基于我们选出的最佳阈值，我们来认为确定y_predict，并确定在这个阈值下的recall和准确度的值:
from sklearn.metrics import accuracy_score as AC times = time() clf = SVC(kernel = "linear",C=3.1663157894736838,cache_size = 5000 ,class_weight = "balanced" ).fit(Xtrain, Ytrain) prob = pd.DataFrame(clf.decision_function(Xtest)) prob.loc[prob.iloc[:,0] >= thresholds[maxindex],"y_pred"]=1 prob.loc[prob.iloc[:,0] < thresholds[maxindex],"y_pred"]=0 prob.loc[:,"y_pred"].isnull().sum() #检查模型本身的准确度 score = AC(Ytest,prob.loc[:,"y_pred"].values) recall = recall_score(Ytest, prob.loc[:,"y_pred"]) print("testing accuracy %f,recall is %f" % (score,recall)) print(datetime.datetime.fromtimestamp(time()-times).strftime("%M:%S:%f"))
6 SVM总结&结语
本文来自博客园，作者：秋华，转载请注明原文链接：https://www.cnblogs.com/qiu-hua/p/14960826.html
