Titanic: Machine Learning from Disaster初练习详解

一、预览数据

#-*-coding=utf-8-*-import pandas as pdimport numpy as npfrom pandas import DataFrameimport matplotlib.pyplot as plt import seaborn as sns#sns.set_style('whitegrid')from sklearn.linear_model import LogisticRegressionfrom sklearn.ensemble import RandomForestClassifiertrain=pd.read_csv(r'C:\Users\0011\Desktop\kaggle\train.csv')#891*12test=pd.read_csv(r'C:\Users\0011\Desktop\kaggle\test.csv')#891*11train.info()train_dms=pd.get_dummies(train['Sex'])#891*2train1=pd.concat([train,train_dms],axis=1)#891*（12+2）

发现age，Cabin, Embarked有缺失值。

看看处理缺失后的年龄和存活的分布

#-*-coding=utf-8-*-import pandas as pdimport numpy as npfrom pandas import DataFrameimport matplotlib.pyplot as plt import seaborn as sns#sns.set_style('whitegrid')from sklearn.linear_model import LogisticRegressionfrom sklearn.ensemble import RandomForestClassifiertrain=pd.read_csv(r'C:\Users\0011\Desktop\kaggle\train.csv')#891*12test=pd.read_csv(r'C:\Users\0011\Desktop\kaggle\test.csv')#891*11train_dms=pd.get_dummies(train['Sex'])#891*2train1=pd.concat([train,train_dms],axis=1)#891*（12+2）nan_num = train['Age'].isnull().sum()# there are 177 missing value, fill with random intage_mean = train['Age'].mean()age_std = train['Age'].std()filling = np.random.randint(age_mean-age_std, age_mean+age_std, size=nan_num)train['Age'][train['Age'].isnull()==True] = fillingnan_num = train['Age'].isnull().sum()# dealing the missing val in testnan_num = test['Age'].isnull().sum()# 86 nullage_mean = test['Age'].mean()age_std = test['Age'].std()filling = np.random.randint(age_mean-age_std,age_mean+age_std,size=nan_num)test['Age'][test['Age'].isnull()==True]=fillingnan_num = test['Age'].isnull().sum()#look into the age cols = sns.FacetGrid(train,hue='Survived',aspect=3)#aspect表示纵横比s.map(sns.kdeplot,'Age',shade=True)#核密度统计方式s.set(xlim=(0,train['Age'].max()))s.add_legend()

由于随机数的影响：

def under15(row):    result = 0.0    if row<15:        result = 1.0    return resultdef young(row):    result = 0.0    if row>=15 and row<30:        result = 1.0    return resulttrain['under15'] = train['Age'].apply(under15)test['under15'] = test['Age'].apply(under15)train['young'] = train['Age'].apply(young)test['young'] = test['Age'].apply(young)

#family# chekprint(train['SibSp'].value_counts(dropna=False))print(train['Parch'].value_counts(dropna=False))sns.factorplot('SibSp','Survived',data=train,size=5)sns.factorplot('Parch','Survived',data=train,size=5)'''through the plot, we suggest that with more family member, the survival rate will drop, we can create the new coladd up the parch and sibsp to check our theory''' train['family'] = train['SibSp'] + train['Parch']test['family'] = test['SibSp'] + test['Parch']sns.factorplot('family','Survived',data=train,size=5)train.drop(['SibSp','Parch'],axis=1,inplace=True)test.drop(['SibSp','Parch'],axis=1,inplace=True)

# fare# checking null, found one in test group. leave it alone til we find outprint(train.Fare.isnull().sum())print(test.Fare.isnull().sum())

result:

0
1

sns.factorplot('Survived','Fare',data=train,size=5)

i=train.Embarked.value_counts()print(i)S    644C    168Q     77Name: Embarked, dtype: int64

二、模型预测

# testing1, using all the featuretrain_ft=train.drop('Survived',axis=1)train_y=train['Survived']#set kfkf = KFold(n_splits=3,random_state=1)acc_lst = []ml(train_ft,train_y,'test_1')

# testing 2, lose youngtrain_ft_2=train.drop(['Survived','young'],axis=1)test_2 = test.drop('young',axis=1)train_ft.head()# mlkf = KFold(n_splits=3,random_state=1)acc_lst=[]ml(train_ft_2,train_y,'test_2')

#test3, lose young, ctrain_ft_3=train.drop(['Survived','young','C'],axis=1)test_3 = test.drop(['young','C'],axis=1)train_ft.head()# mlkf = KFold(n_splits=3,random_state=1)acc_lst = []ml(train_ft_3,train_y,'test_3')

# test4, no FAREtrain_ft_4=train.drop(['Survived','Fare'],axis=1)test_4 = test.drop(['Fare'],axis=1)train_ft.head()# mlkf = KFold(n_splits=3,random_state=1)acc_lst = []ml(train_ft_4,train_y,'test_4')

# test5, get rid of c train_ft_5=train.drop(['Survived','C'],axis=1)test_5 = test.drop('C',axis=1)# mlkf = KFold(n_splits=3,random_state=1)acc_lst = []ml(train_ft_5,train_y,'test_5')

# test6, lose Fare and youngtrain_ft_6=train.drop(['Survived','Fare','young'],axis=1)test_6 = test.drop(['Fare','young'],axis=1)train_ft.head()# mlkf = KFold(n_splits=3,random_state=1)acc_lst = []ml(train_ft_6,train_y,'test_6')

accuracy_df=pd.DataFrame(data=accuracy,                         index=['test1','test2','test3','test4','test5','test6'],                         columns=['logistic','rf','svc','knn'])

原作者最好是test4

svc 0.832727但是提交后结果：

总结一下源码：

#-*-coding=utf-8-*-import pandas as pdpd.options.mode.chained_assignment = Noneimport numpy as npfrom pandas import DataFrameimport matplotlib.pyplot as plt import seaborn as snssns.set_style('whitegrid')from sklearn.linear_model import LogisticRegressionfrom sklearn.ensemble import RandomForestClassifierfrom sklearn.svm import SVC, LinearSVCfrom sklearn.ensemble import GradientBoostingClassifierfrom sklearn.neighbors import KNeighborsClassifierfrom sklearn.naive_bayes import GaussianNBfrom sklearn.metrics import accuracy_scorefrom sklearn.model_selection import cross_val_score, KFold#第一部分. Exploring the Datatrain=pd.read_csv(r'C:\Users\0011\Desktop\kaggle\train.csv')test=pd.read_csv(r'C:\Users\0011\Desktop\kaggle\test.csv')#print(train.describe())  #train.info()#sns.factorplot("Pclass","Survived",data=train)#plt.show()#2. Data Cleaning and Features Choosingdef dummies(col,train,test):        #定义指示函数处理某个离散型特征,等价把各个变量用数字表示1,2,3    train_dms=pd.get_dummies(train[col])#891*sex(2),Pclass(3)    test_dms =pd.get_dummies(test[col])    train=pd.concat([train,train_dms],axis=1)#axis=1表示第一行#891*（12+2）    test =pd.concat([test,test_dms],axis=1)    train.drop(col,axis=1,inplace=True)#等价于train=train.drop(col,axis=1,inplace=False)    test.drop(col,axis=1,inplace=True)    return train, test#get rid of uselessdropping = ['PassengerId', 'Name', 'Ticket']train.drop(dropping,axis=1, inplace=True)test.drop(dropping,axis=1, inplace=True)#print(train.Pclass.value_counts())#Returns object containing counts of unique values.#pclasstrain, test = dummies('Pclass',train,test)'''观察Sex和Survived的关系，女性生还率显著高于男性'''#print(train.Sex.value_counts(dropna=False))#sns.factorplot('Sex','Survived',data=train)#plt.show()train,test = dummies('Sex',train,test)train.drop('male',axis=1,inplace=True)test.drop('male',axis=1,inplace=True)# ensure no na contained#age #dealing the missing datanan_num = train['Age'].isnull().sum()# there are 177 missing value, fill with random intage_mean = train['Age'].mean()#平均值age_std = train['Age'].std()#方差filling = np.random.randint(age_mean-age_std, age_mean+age_std, size=nan_num)train['Age'][train['Age'].isnull()==True] = fillingnan_num = train['Age'].isnull().sum()# dealing the missing val in testnan_num = test['Age'].isnull().sum()age_mean=test['Age'].mean()age_std=test['Age'].std()filling=np.random.randint(age_mean-age_std, age_mean+age_std, size=nan_num)test['Age'][test['Age'].isnull()==True]=fillingnan_num = test['Age'].isnull().sum()# from the graph, we see that the survival rate of children# is higher than other and the 15-30 survival rate is lowerdef under15(row):    result = 0.0    if row<15:        result = 1.0    return resultdef young(row):    result = 0.0    if row>=15 and row<30:        result = 1.0    return resulttrain['under15'] = train['Age'].apply(under15)test['under15'] = test['Age'].apply(under15)train['young'] = train['Age'].apply(young)test['young'] = test['Age'].apply(young)train.drop('Age',axis=1,inplace=True)test.drop('Age',axis=1,inplace=True)#family cheak can adjust(i think)train['family'] = train['SibSp'] + train['Parch']test['family'] = test['SibSp'] + test['Parch']train.drop(['SibSp','Parch'],axis=1,inplace=True)test.drop(['SibSp','Parch'],axis=1,inplace=True)# fare# checking null, found one in test group. leave it alone til we find outtest['Fare'].fillna(test['Fare'].median(),inplace=True)#Cabin# checking missing val# 687 out of 891 are missing, drop this coltrain.drop('Cabin',axis=1,inplace=True)test.drop('Cabin',axis=1,inplace=True)#Embark#train.Embarked.isnull().sum()# 2 missing value# fill the majority val,'s', into missing val coltrain['Embarked'].fillna('S',inplace=True)# c has higher survival rate, drop the other twotrain,test = dummies('Embarked',train,test)train.drop(['S','Q'],axis=1,inplace=True)test.drop(['S','Q'],axis=1,inplace=True)#第二部分kf = KFold(n_splits=3,random_state=1)def modeling(clf,ft,target):    acc = cross_val_score(clf,ft,target,cv=kf)#cross_val_score(clf,X,y)    acc_lst.append(acc.mean())    return accuracy = []def ml(ft,target,time):    accuracy.append(acc_lst)    #logisticregression    logreg = LogisticRegression()    modeling(logreg,ft,target)    #RandomForest    rf = RandomForestClassifier(n_estimators=50,min_samples_split=4,min_samples_leaf=2)    modeling(rf,ft,target)    #svc    svc = SVC()    modeling(svc,ft,target)    #knn    knn = KNeighborsClassifier(n_neighbors = 3)    modeling(knn,ft,target)            # see the coefficient    logreg.fit(ft,target)    feature = pd.DataFrame(ft.columns)    feature.columns = ['Features']    feature["Coefficient Estimate"] = pd.Series(logreg.coef_[0])    #print(feature)    return # testing no.1, using all the featuretrain_ft=train.drop('Survived',axis=1)train_y=train['Survived']#set kfacc_lst = []ml(train_ft,train_y,'test_1')# testing 2, lose youngtrain_ft_2=train.drop(['Survived','young'],axis=1)test_2 = test.drop('young',axis=1)train_ft.head()# mlacc_lst=[]ml(train_ft_2,train_y,'test_2')#test3, lose young, ctrain_ft_3=train.drop(['Survived','young','C'],axis=1)test_3 = test.drop(['young','C'],axis=1)train_ft.head()# mlacc_lst = []ml(train_ft_3,train_y,'test_3')# test4, no FAREtrain_ft_4=train.drop(['Survived','Fare'],axis=1)test_4 = test.drop(['Fare'],axis=1)train_ft.head()# mlacc_lst = []ml(train_ft_4,train_y,'test_4')# test5, get rid of c train_ft_5=train.drop(['Survived','C'],axis=1)test_5 = test.drop('C',axis=1)# mlacc_lst = []ml(train_ft_5,train_y,'test_5')# test6, lose Fare and youngtrain_ft_6=train.drop(['Survived','Fare','young'],axis=1)test_6 = test.drop(['Fare','young'],axis=1)train_ft.head()# mlacc_lst = []ml(train_ft_6,train_y,'test_6')accuracy_df=pd.DataFrame(data=accuracy,                         index=['test1','test2','test3','test4','test5','test6'],                         columns=['logistic','rf','svc','knn'])#test4 svc as submissionsvc = SVC()svc.fit(train_ft_4,train_y)svc_pred = svc.predict(test_4)print(svc.score(train_ft_4,train_y))test=pd.read_csv(r'C:\Users\0011\Desktop\kaggle\test.csv')submission = pd.DataFrame({        'PassengerId': test['PassengerId'],        'Survived': svc_pred    })#submission.to_csv("kaggle.csv", index=False)

三、试试自己的改进

#自己添加了一些都试试不同组合，没有什么提升效果def up60(row):    result = 0.0    if row>60:        result = 1.0    return resultdef f38t48(row):    result = 0.0    if row>=38 and row<48:        result = 1.0    return resultdef f48t60(row):    result = 0.0    if row>=48 and row<60:        result = 1.0    return resulttrain['up60'] = train['Age'].apply(up60)test['up60'] = test['Age'].apply(up60)train['f38t48'] = train['Age'].apply(f38t48)test['f38t48'] = test['Age'].apply(f38t48)train['f48t60'] = train['Age'].apply(f48t60)test['f48t60'] = test['Age'].apply(f48t60)

试试姓氏和年龄的关系结果提交的过拟合了

# ensure no na contained#age train['Initial']=train.Name.str.extract('([A-Za-z]+)\.')test['Initial']=test.Name.str.extract('([A-Za-z]+)\.')train['Initial'].replace(['Mlle','Mme','Ms','Dr','Major','Lady','Countess',     'Jonkheer','Col','Rev',     'Capt','Sir','Don'],['Miss','Miss','Miss','Mr','Mr','Mrs','Mrs',    'Other','Other','Other','Mr','Mr','Mr'],inplace=True)        test['Initial'].replace(['Mlle','Mme','Ms','Dr','Major','Lady','Countess',     'Jonkheer','Col','Rev',     'Capt','Sir','Don'],['Miss','Miss','Miss','Mr','Mr','Mrs','Mrs',    'Other','Other','Other','Mr','Mr','Mr'],inplace=True)#检查每一个姓氏的平均年龄#print(train.groupby('Initial')['Age'].mean()) ## Assigning the NaN Values with the Ceil values of the mean agestrain.loc[(train.Age.isnull())&(train.Initial=='Mr'),'Age']=33train.loc[(train.Age.isnull())&(train.Initial=='Mrs'),'Age']=36train.loc[(train.Age.isnull())&(train.Initial=='Master'),'Age']=5train.loc[(train.Age.isnull())&(train.Initial=='Miss'),'Age']=22train.loc[(train.Age.isnull())&(train.Initial=='Other'),'Age']=46test.loc[(test.Age.isnull())&(test.Initial=='Mr'),'Age']=33test.loc[(test.Age.isnull())&(test.Initial=='Mrs'),'Age']=36test.loc[(test.Age.isnull())&(test.Initial=='Master'),'Age']=5test.loc[(test.Age.isnull())&(test.Initial=='Miss'),'Age']=22test.loc[(test.Age.isnull())&(test.Initial=='Other'),'Age']=46train.drop(['Name','Initial'],axis=1, inplace=True)test.drop(['Name','Initial'],axis=1, inplace=True)

还不错有进步：

更多数据分析图参看

参考