泰坦尼克号生存预测(一)-- 数据处理

项目及数据集来自Kaggle。

持续更新中......

1. 提出问题

建立模型预测乘客是否生还。

2. 理解数据

数据特征含义:survival为目标变量,其他为特征。

Variable Definition Key
survival Survival 0 = No, 1 = Yes
pclass Ticket class 1 = 1st, 2 = 2nd, 3 = 3rd
sex Sex  
Age Age in years  
sibsp # of siblings / spouses aboard the Titanic  
parch # of parents / children aboard the Titanic  
ticket Ticket number  
fare Passenger fare  
cabin Cabin number  
embarked Port of Embarkation C = Cherbourg, Q = Queenstown, S = Southampton
# load libraries of anlysis and visualization
import numpy as np
import pandas as pd
import re  # Regular Expression operations
import matplotlib.pyplot as plt
%matplotlib inline

train = pd.read_csv('train.csv')
test = pd.read_csv('test.csv')

# 观察数据
train.head()  #前5行数据
train.sample(5)  #随机5行数据
train.describe()  #各列统计数据
train.dtypes  #数据类型

# 载入seaborn作图
import seaborn as sns

f,ax = plt.subplots(3,4,figsize=(20,16))
sns.countplot('Pclass',data=train,ax=ax[0,0])
sns.countplot('Sex',data=train,ax=ax[0,1])
sns.boxplot(x='Pclass',y='Age',data=train,ax=ax[0,2])
sns.distplot(train['Fare'].dropna(),ax=ax[2,0],kde=False,color='b')
sns.countplot('Embarked',data=train,ax=ax[2,2])

sns.countplot('SibSp',hue='Survived',data=train,ax=ax[0,3],palette='husl')
sns.countplot('Parch',hue='Survived',data=train,ax=ax[1,3],palette='husl')
sns.countplot('Embarked',hue='Survived',data=train,ax=ax[2,3],palette='husl')
sns.countplot('Pclass',hue='Survived',data=train,ax=ax[1,0],palette='husl')
sns.countplot('Sex',hue='Survived',data=train,ax=ax[1,1],palette='husl')

sns.distplot(train[train['Survived']==0]['Age'].dropna(),ax=ax[1,2],kde=False,color='r',bins=5)
sns.distplot(train[train['Survived']==1]['Age'].dropna(),ax=ax[1,2],kde=False,color='g',bins=5)

sns.swarmplot(x='Pclass',y='Fare',hue='Survived',data=train,ax=ax[2,1],palette='husl')

ax[0,0].set_title('Total Passengers by Class')
ax[0,1].set_title('Total Passengers by Sex')
ax[0,2].set_title('Age boxplot by Class')
ax[0,3].set_title('Survival Rate by SibSp')
ax[1,0].set_title('Survival Rate by Pclass')
ax[1,1].set_title('Survival Rate by gender')
ax[1,2].set_title('Survival Rate by Age')
ax[1,3].set_title('Survival Rate by Parch')
ax[2,0].set_title('Fare Distribution')
ax[2,1].set_title('Survival Rate by Fare and Pclass')
ax[2,2].set_title('Total Passengers by Embarked')
ax[2,3].set_title('Survival Rate by Embarked')

3. 数据清理

a. 找出异常值和离群点

# 检测异常值  因为此数据中没有明显异常点,故检测离群点(1.5个IQR(四分位距)以外的点)
'''
定义离群点函数
输入:dataset,MAX离群特征个数n,特征名
输出:超过n个离群特征的样本index
'''
# 调取collections的Counter,用于对list计数
from collections import Counter

def detect_outliers(df,n,feature):
    outlier_indices = []
    for f in feature:
        # 1st quartile(25%)
        Q1=np.percentile(df[f],25)
        # 3rd quartile(75%)
        Q3=np.percentile(df[f],75)
        # Interquartile range四分位距
        IQR=Q3-Q1
        outlier_step=1.5*IQR
        # 生成该特征中为离群点的样本index
        outlier_list_col = df[(df[f]<(Q1-outlier_step))|(df[f]>(Q3+outlier_step))].index
        # 生成一个含有离群点样本index的list
        outlier_indices.extend(outlier_list_col)
    # 对该list计数,生成字典key为要计数的值,value为key的计数值
    outlier_indices = Counter(outlier_indices)
    multiple_outliers = list(k for k,v in outlier_indices.items() if v>n)
    return multiple_outliers

# 找出同时有两个以上特征为离群点的样本
Outliers_to_drop=detect_outliers(train,2,['Fare','Age','SibSp','Parch'])
train.loc[Outliers_to_drop]

由上述结果可以: 1. 存在10个以上超过2个特征为离群点的样本; 2. 有三位乘客费付了较高的票价263;3. 有7位乘客有较多的兄弟姐妹或伴侣在此船上;4. 结合之前train.describe()结果,这些离群点并不是偏得离谱,因此,应该保留下来。

b. 缺失值处理

# 统计各列缺失值个数
train.isnull().sum()
test.isnull().sum()

在train和test两个数据集中,"Age", "Cabin"中含有较多缺失值,在train数据集中embarked含有两个缺失值,在test数据集中Fare有一个缺失值。

4. 特征工程-二变量统计分析

对特征分析第一步就要看特征是numerical/ordinal values,一般分两种情况:

1) 定性型Qualitative data: discrete离散型

名义型数据如姓名,或分类型数据如性别等

2) Numeric/Qualitative data连续型或可排序的数据:

数值离散型数据:如Pcalss等级这类可以排序的数据,或连续型数据

a. 上船地点

# Embarked
sns.barplot(x = 'Embarked',y='Survived',hue='Sex',data=train)

单从此图看,女性的存活率明显高于男性,从哪里上船与存活率没有明显的关系,需要进一步探索。

b. 姓名

姓名长度

train['Name_length']=train['Name'].apply(len)
sum_name = train[['Name_length','Survived']].groupby(['Name_length'],as_index=False).sum()
average_name = train[['Name_length','Survived']].groupby(['Name_length'],as_index=False).mean()
f,(axis1,axis2,axis3) = plt.subplots(3,1,figsize=(18,6))
sns.barplot(x = 'Name_length',y='Survived',data=sum_name,ax=axis1)
sns.barplot(x='Name_length',y='Survived',data=average_name,ax=axis2)
sns.pointplot(x='Name_length',y='Survived',data=train,ax=axis3)

c. 性别

将性别转换为数值型二分类数据

full_data = [train,test]
Survival = train['Survived']

# 将性别数值化
for dataset in full_data:
    dataset['Sex']=dataset['Sex'].map({'female':0,'male':1}).astype(int)

d. 年龄

# Age
# 不同年龄存活率分布
# sns.FacetGrid可以绘制多个轴变量相同的图,aspect设置图aspect ratio纵横比,size设置图片高度,通常需要用.map()作图。
a=sns.FacetGrid(train,hue='Survived',aspect=6)
a.map(sns.kdeplot,'Age',shade=True)
a.set(xlim=[0,train['Age'].max()])
a.add_legend()

由以上核密度图可知:存活的人中,小孩和30左右的人占比较高,而未存活的人集中在23左右,且右偏。

将年龄离散化:

# 由于年龄有较多缺失值,用均值正负一个标准差内的值对其填充
for dataset in full_data:
    age_avg = dataset['Age'].mean()
    age_std = dataset['Age'].std()
    age_null_count = dataset['Age'].isnull().sum()
    # 对缺失值填充
    age_null_random_list = np.random.randint(age_avg-age_std,age_avg+age_std,size=age_null_count)
    dataset['Age'][np.isnan(dataset['Age'])]=age_null_random_list
    dataset['Age']=dataset['Age'].astype(int)
    # 对年龄离散化
    # 若使用pd.qcut()可自动离散化,dataset['Age]=pd.qcout(dataset['Age'],6,labels=False)
    # 本例自定义
    dataset.loc[dataset['Age']<=14,'Age_level'] = 1
    dataset.loc[(dataset['Age']>14)&(dataset['Age']<=30),'Age_level'] = 2
    dataset.loc[(dataset['Age']>30)&(dataset['Age']<=40),'Age_level'] = 3
    dataset.loc[(dataset['Age']>40)&(dataset['Age']<=50),'Age_level'] = 4
    dataset.loc[(dataset['Age']>50)&(dataset['Age']<=60),'Age_level'] = 5
    dataset.loc[(dataset['Age']>60),'Age_level'] = 6

train['Age_level'].value_counts()
Age_Surv=train[['Age_level','Survived']].groupby('Age_level',as_index=False).mean().sort_values(by='Age_level',ascending=False)
sns.barplot(x='Age_level',y='Survived',data=Age_Surv)

由上图可知:年龄在14岁以下的儿童存活率明显高于其他年龄段的人,年龄在60岁以上的老年人存活率明显低于其他年龄段的人。

e. 家庭:SibSp/Parch

# 用 familysize 来表示SibSp和Parch
for dataset in full_data:
    dataset['Familysize']=dataset['SibSp']+dataset['Parch']+1
    # 用isAlone列表示是否独自一人上船
    dataset['isAlone']=0
    dataset.loc[dataset['Familysize']==1,'isAlone']=1
    # 用boy列表示是否为男孩
    dataset['boy']=0
    dataset.loc[(dataset['Sex']==1)&(dataset['Age_level']==0),'boy']=1

f, (axis1,axis2) = plt.subplots(1,2, figsize=(18,6))
sns.barplot(x='Familysize',y='Survived',hue='Sex',data=train,ax=axis1)
sns.barplot(x='isAlone',y='Survived',hue='Sex',data=train,ax=axis2)

由上图可知,是否有家人陪同isAlone存活率没有明显的差别,反而是男女的存活率有明显的差别。

f. 费用

# 使用cufflinks画交互图
import cufflinks as cf
cf.go_offline()
train['Fare'].iplot(kind=hist,bins=30)

交互式图表的优点是鼠标指哪里就显示哪里的数值。

# 处理【'Fare'】缺失值,用中位数填充
for dataset in full_data:
    dataset['Fare'].fillna(dataset['Fare'].median)

# 看看费用分布图
sns.distplot(train['Fare'],color = 'm',label='Skewness: %.2f'%(train['Fare'].skew()))
plt.legend(loc='best')

由以上费用分布图可以看出,费用分布严重右偏,这会造成统计量右偏,因此,可以将数据对数化来降低偏度。

# 取对数,np.log()
for dataset in full_data:
    dataset['Fare_log']=dataset['Fare'].map(lambda i: np.log(i) if i > 0 else 0)

f, ax = plt.subplots(figsize=(20,6))
sns.distplot(train['Fare_log'][train['Survived']==0],color='r',label='Skewness: %.2f'%(train['Fare_log'].skew()))
sns.distplot(train['Fare_log'][train['Survived']==1],color='b',label='Skewness: %.2f'%(train['Fare_log'].skew()))
plt.legend(['Not survived','Survived'])

由上图可以看出,log['Fare'] 高于2.7的有更高的存活率,而低于2.7的存活率较低。

将数值离散化:

# 将log('Fare')离散化分析
for dataset in full_data:
    dataset.loc[dataset['Fare_log']< 2.7,'Fare_log'] = 1
    dataset.loc[(dataset['Fare_log']>= 2.7)&(dataset['Fare_log']< 3.2),'Fare_log'] = 2
    dataset.loc[dataset['Fare_log']>=3.2,'Fare_log'] = 3
train['Fare_log'].value_counts()

g. 客舱

有无客舱

# 若无客舱标为0,有客舱标记为1
for dataset in full_data:
    dataset['Has_Cabin'] = dataset['Cabin'].apply(lambda i: 0 if type(i)==float else 1)
train[['Has_Cabin','Survived']].groupby(['Has_Cabin'], as_index=False).sum().sort_values(by='Survived',ascending=False)

train[['Has_Cabin','Survived']].groupby(['Has_Cabin'], as_index=False).mean().sort_values(by='Survived',ascending=False)
  Has_Cabin Survived
0 0 206
1 1 136
  Has_Cabin Survived
1 1 0.666667
0 0 0.299854

拥有客舱的乘客存活率明显高于没有客舱的乘客。

客舱分类和数值化:

Cabin里的C123此处的C代表甲板等级Deck:

# U代表Unknown没有cabin的
deck = {'A':1,'B':2,'C':3,'D':4,'E':5,'F':6,'G':7,'U':8}
for dataset in full_data:
    dataset['Cabin']=dataset['Cabin'].fillna('U')
    dataset['Deck']=dataset['Cabin'].map(lambda x: re.compile('[A-Za-z]+').search(x).group())
    dataset['Deck']=dataset['Deck'].map(deck)
    dataset['Deck']=dataset['Deck'].fillna(0)
    dataset['Deck']=dataset['Deck'].astype(int)

train['Deck'].value_counts()

sns.barplot(x='Deck',y='Survived',data=train,order=[1,2,3,4,5,6,7,8])
8    687
3     59
2     47
4     33
5     32
1     15
6     13
7      4
0      1

由以上图表,将Deck分为三类:

for dataset in full_data:
    dataset.loc[(dataset['Deck']<=1),'Deck']=1
    dataset.loc[(dataset['Deck']>1)&(dataset['Deck']<7),'Deck']=0
    dataset.loc[(dataset['Deck']>=7),'Deck']=2
    
train[['Deck','Survived']].groupby('Deck',as_index=False).mean().sort_values(by='Survived',ascending=False)
  Deck Survived
0 0 0.690217
1 1 0.437500
2 2 0.301013

h. 上船地点

for dataset in full_data:
    # 缺失值填充为'S'
    dataset['Embarked']=dataset['Embarked'].fillna('S')
    # 数值化 
    dataset['Embarked']=dataset['Embarked'].map({'S': 1, 'C':2, 'Q':3}).astype(int)

train_pivot = pd.pivot_table(train, index='Embarked', columns = 'Pclass', values = 'Survived', aggfunc=np.mean, margins=True)
def train_nagetive_red(val):
    color = 'red' if val < 0.4 else 'black'
    return 'color: %s'%color
train_pivot = train_pivot.style.applymap(train_nagetive_red)
train_pivot

可以看到,无论从哪里上船,三等舱的乘客生存率都小于0.4;此外,若不考虑客舱等级,从S和Q上船的乘客生存率要小于从C上船的乘客。

i. 乘客的title

# 用正则表达式寻找Name里的Title
def get_title(name):
    # Mr.  Miss.等
    title_search = re.search('([A-Za-z]+)\.',name)
    if title_search:
        # 返回找到的第一个字符串group(1)
        return title_search.group(1)
    return ""

for dataset in full_data:
    dataset['Title']=dataset['Name'].apply(get_title)

f, ax = plt.subplots(1,figsize=(18,6))
sns.barplot(x='Title',y='Survived',data=train,ax=ax)

由上图可以看到不同Title之间生存率有明显差别,其中,Mme,Ms,Lady,Sir,Mlle,Countess的生存率为100%,但是单从上图中无法得出这些人的生存几率就是高于其他人,因为很可能这些人的样本数非常少,是个例。因此,需要结合总人数来看:

A = train[['Title','Survived']].groupby('Title').mean()
B = train[['Title','Survived']].groupby('Title').count()
C = pd.concat([A,B],axis=1)
C.columns = ['Survive rate','Total']
f,ax = plt.subplots(figsize=(18,6))
C.plot(kind='bar',x=C.index,y='Survive rate',ax=ax)
C.plot(kind='line',x=C.index,y='Total',secondary_y=True,ax=ax)

  Survive rate Total
Title    
Capt 0.000000 1
Col 0.500000 2
Countess 1.000000 1
Don 0.000000 1
Dr 0.428571 7
Jonkheer 0.000000 1
Lady 1.000000 1
Major 0.500000 2
Master 0.575000 40
Miss 0.697802 182
Mlle 1.000000 2
Mme 1.000000 1
Mr 0.156673 517
Mrs 0.792000 125
Ms 1.000000 1
Rev 0.000000 6
Sir 1.000000 1

结合总人数来看生存率,这些生存率为100%的Title总人数都不多于2人,当然,也有些人数和生存率均很高的Title,例如,Mrs/Miss/Master,而总体来看,Mr的生存率非常低。

要使用监督学习模型,需对Title进行分类和数值化:

  1. Mme, Ms, Lady, Sir, Mlle, Countess: 100%.
  2. Mrs, Miss: around 70% survival
  3. Master: around 60%
  4. Dr, Major, Col: around 40%
  5. Mr: below 20%
  6. Don, Rev, Capt, Jonkheer: 0
for dataset in full_data:
    dataset['Title']=dataset['Title'].replace(['Mme', 'Ms', 'Lady', 'Sir', 'Mlle', 'Countess'],'MMLSMC')
    dataset['Title'] = dataset['Title'].replace(['Mrs', 'Miss'], 'MM')
    dataset['Title'] = dataset['Title'].replace(['Dr', 'Major', 'Col'], 'DMC')
    dataset['Title'] = dataset['Title'].replace(['Don', 'Rev', 'Capt', 'Jonkheer'],'DRCJ')
    # title map
    title_mapping = {"MMLSMC":1,"MM":2,"Master":3,"DMC":4,"Mr": 5,"DRCJ":6}
    dataset['Title']=dataset['Title'].map(title_mapping)

# test中有一个“Dona.”没有赋值,取中间的值赋予它
test['Title']=test['Title'].fillna(4)

train[['Title','Survived']].groupby(['Title'],as_index=False).mean().sort_values(by='Survived',ascending=False)
  Title Survived
0 1 1.000000
1 2 0.736156
2 3 0.575000
3 4 0.454545
4 5 0.156673
5 6 0.000000

猜你喜欢

转载自blog.csdn.net/huangxiaoyun1900/article/details/82426132
今日推荐