泰坦尼克号预测生还案例

一、背景

 Titanic: Machine Learning from Disaster-https://www.kaggle.com/c/titanic/data,必须先登录kaggle

  • 就是那个大家都熟悉的『Jack and Rose』的故事,豪华游艇倒了,大家都惊恐逃生,可是救生艇的数量有限,无法人人都有,副船长发话了『 lady and kid first!』,所以是否获救其实并非随机,而是基于一些背景有rank先后的。

  • 训练和测试数据是一些乘客的个人信息以及存活状况,要尝试根据它生成合适的模型并预测其他人的存活状况。

 

 观察train数据发现总共有12列,其中Survived字段表示的是该乘客是否获救,其余都是乘客的个人信息,包括:  

  • PassengerId => 乘客ID
  • Pclass => 乘客等级(1/2/3等舱位),1st-Upper,2nd-Middle,3rd-Lower
  • Name => 乘客姓名
  • Sex => 性别
  • Age => 年龄
  • SibSp => 同船的堂兄弟/妹个数
  • Parch => 同船的父母与小孩个数
  • Ticket => Ticket number
  • Fare => 票价
  • Cabin => Cabin number
  • Embarked => 登船港口:C = Cherbourg, Q = Queenstown, S = Southampton

 其中测试集示例如下:

 

二、代码分析-参考别人的决策树相关算法——XGBoost原理分析及实例实现(三)

 

 2.1 查看数据的整体信息

  可以看看哪些各个列变量的缺失值情况,比如Age,Cabin

import pandas as pd
import numpy as np

train_file = "train.csv"
test_file = "test.csv"
test_result_file = "gender_submission.csv"

train = pd.read_csv(train_file)
test = pd.read_csv(test_file)
test_y = pd.read_csv(test_result_file)#test也有标签,用于核对模型对test数据预测的结果好坏

full_data = [train,test]
print(train.info())
'''
<class 'pandas.core.frame.DataFrame'>
RangeIndex:      891 entries, 0 to 890
Data columns     (total 12 columns):
PassengerId          891 non-null int64
Survived               891 non-null int64
Pclass                 891 non-null int64
Name                   891 non-null object
Sex                    891 non-null object
Age                      714 non-null float64
SibSp                  891 non-null int64
Parch                  891 non-null int64
Ticket                 891 non-null object
Fare                   891 non-null float64
Cabin                   204 non-null object
Embarked               889 non-null object
dtypes: float64(2), int64(5), object(5)
memory usage: 83.6+ KB
'''         

 

 2.2 根据train.csv中各个Variable的取值和特性进行数据处理

  主要查看数据的各个Variable对Survived的影响来确定是否该Variable对生还有影响

 

  1.PassengerId 和 Survived:PassengerId是各个乘客的ID,每个ID号各不相同,基本没有什么数据挖掘意义,对需要预测的存活性几乎没有影响。

  2.Pclass:为船票类型,离散数据(不需要进行特别处理),没有缺失值。该变量的取值情况如下

###### in
print (train['Pclass'].value_counts(sort=False).sort_index())
###### out
1    216
2    184
3    491
###### Pclass和Survived的影响
#计算出每个Pclass属性的取值中存活的人的比例
print train[['Pclass','Survived']].groupby('Pclass',as_index=False).mean()
#巧妙的利用groupby().mean()函数,如Pclass == 1中有4个1(存活),1个0(死亡),则mean()后4个1的和/5个 = 0.8
###### out
   Pclass  Survived
0       1  0.629630
1       2  0.472826
2       3  0.242363

  从输出的生还率可以看出,不同的Pclass类型对生还率影响还是很大的,所以选取该属性作为最终的模型的特征之一,取值为1,2,3

 

  3.Sex 性别,连续型数据特征,没有缺失值。该变量的取值情况如下

###### in
print (train['Sex'].value_counts(sort=False).sort_index())
###### out
female    314
male      577
###### Sex和Survived的影响
#计算出每个Sex属性的取值中存活的人的比例
print train[['Sex','Survived']].groupby('Sex',as_index=False).mean()
###### out
      Sex  Survived
0  female  0.742038
1    male  0.188908

  从输出的生还率可以看出,不同的Sex类型对生还率影响还是很大的,所以选取该属性作为最终的模型.对于字符串数据特征值的处理,可以将两个字符串值映射到两个数值0,1上。

 

  4.Age 年龄是连续型数据,该属性包含较多的缺失值,不宜删除缺失值所在的行的数据记录。此处不仅需要对缺失值进行处理,而且需要对该连续型数据进行处理。

  对于该属性的缺失值处理:

    方法一,默认填充值的范围[(mean - std) ,(mean + std)]。

    方法二,将缺失的Age当做label,将其他列的属性当做特征,通过已有的Age的记录训练模型,来预测缺失的Age值。

 

  对该连续型数据进行处理:常用的方法有两种:

    方法一,等距离划分。

    方法二,通过卡方检验/信息增益/GINI系数寻找差异较大的分裂点

###对于该属性的缺失值处理方式一,方式二在最终的代码仓库中
for dataset in full_data:
    age_avg = dataset['Age'].mean()
    age_std = dataset['Age'].std()
    
    age_null_count = dataset['Age'].isnull().sum()
    age_default_list = np.random.randint(low=age_avg-age_std,high=age_avg+age_std,size=age_null_count)
    
    dataset['Age'][np.isnan(dataset['Age'])] = age_default_list
    dataset['Age'] = dataset['Age'].astype(int)

###对该连续型数据进行处理方式二
train['CategoricalAge'] = pd.cut(train['Age'], 5)
print (train[['CategoricalAge', 'Survived']].groupby(['CategoricalAge'], as_index=False).mean())
###### out
  CategoricalAge  Survived
0  (-0.08, 16.0]  0.532710
1   (16.0, 32.0]  0.360802
2   (32.0, 48.0]  0.360784
3   (48.0, 64.0]  0.434783
4   (64.0, 80.0]  0.090909

  可以看出对连续型特征Age离散化处理后,各个年龄阶段的存活率还是有差异的,所以可以选取CategoricalAge作为最终模型的一个特征

   

  5.SibSp and Parch:SibSp和Parch分别为同船的兄弟姐妹和父母子女数,离散数据,没有缺失值。于是可以根据该人的家庭情况组合出不同的特征

###### SibSp对Survived的影响
print train[['SibSp','Survived']].groupby('SibSp',as_index=False).mean()
###### Parch对Survived的影响
print train[['Parch','Survived']].groupby('Parch',as_index=False).mean()
###### Parch和SibSp组合对Survived的影响
for dataset in full_data:
    dataset['FamilySize'] = dataset['SibSp'] + dataset['Parch'] + 1
print (train[['FamilySize', 'Survived']].groupby(['FamilySize'], as_index=False).mean())
###### 是否为一个人IsAlone对Survived的影响
train['IsAlone'] = 0
train.loc[train['FamilySize']==1,'IsAlone'] = 1
#或者通过下面的代码来个Alone赋值为1
train['IsAlone'][train['FamilySize'] == 1] = 1
print (train[['IsAlone', 'Survived']].groupby(['IsAlone'],
as_index=False).mean())
###### out 1
   SibSp  Survived 
0      0  0.345395
1      1  0.535885
2      2  0.464286
3      3  0.250000
4      4  0.166667
5      5  0.000000
6      8  0.000000
###### out 2
   Parch  Survived
0      0  0.343658
1      1  0.550847
2      2  0.500000
3      3  0.600000
4      4  0.000000
5      5  0.200000
6      6  0.000000
###### out 3
0           1  0.303538
1           2  0.552795
2           3  0.578431
3           4  0.724138
4           5  0.200000
5           6  0.136364
6           7  0.333333
7           8  0.000000
8          11  0.000000
###### out 4
   IsAlone  Survived
0        0  0.505650
1        1  0.303538

  从输出的生还率可以看出,可以选取的模型特征有Parch和SibSp组合特征FamilySize,Parch,SibSp,IsAlone该四个特征的取值都为离散值

 

  6.Ticket和Cabin:Ticket为船票号码,每个ID的船票号不同,难以进行数据挖掘,所以该列可以舍弃。Cabin为客舱号码,并且对于891条数据记录来说,其缺失值较多,缺失巨大,难以进行填充或者说进行缺失值补充带来的噪音将更多,所以考虑放弃该列

  7.Fare:Fare为船票售价,连续型数据,没有缺失值,需要对该属性值进行离散化处理

for dataset in full_data:
    dataset['Fare'] = dataset['Fare'].fillna(train['Fare'].median())
train['CategoricalFare'] = pd.qcut(train['Fare'],6)
print (train[['CategoricalFare', 'Survived']].groupby(['CategoricalFare'], as_index=False).mean())
###### out
     CategoricalFare  Survived
0    (-0.001, 7.775]  0.205128
1     (7.775, 8.662]  0.190789
2    (8.662, 14.454]  0.366906
3     (14.454, 26.0]  0.436242
4     (26.0, 52.369]  0.417808
5  (52.369, 512.329]  0.697987

  可以看出对连续型特征Fare离散化处理后,各个票价阶段的存活率还是有差异的,所以可以选取CategoricalFare作为最终模型的一个特征。此时分为了6个等样本数阶段

 

  8.Embarked:Embarked是终点城市,字符串型特征值,缺失数极小,所以这里考虑使用该属性最多的值填充

print (train['Embarked'].value_counts(sort=False).sort_index())
###### out
C    168
Q     77
S    644
Name: Embarked, dtype: int64
#### 填充和探索Embarked对Survived的影响
for data in full_data:
    data['Embarked'] = data['Embarked'].fillna('S')
print (train['Embarked'].value_counts(sort=False).sort_index())
print (train[['Embarked', 'Survived']].groupby(['Embarked'], as_index=False).mean())
###### out1
C    168
Q     77
S    646
Name: Embarked, dtype: int64
  Embarked  Survived
0        C  0.553571
1        Q  0.389610
2        S  0.339009

  可以看出不同的Embarked类型对存活率的影响有差异,所以可以选择该列作为最终模型的特征,由于该属性的值是字符型,还需要进行映射处理或者one-hot处理。

 

  9.Name:Name为姓名,字符型特征值,没有缺失值,需要对字符型特征值进行处理。但是观察到Name的取值都是不相同,但其中发现Name的title name是存在类别的关系的。于是可以对Name进行提取出称呼这一类别title name

import re
def get_title_name(name):
    title_s = re.search(' ([A-Za-z]+)\.', name)
    if title_s:
        return title_s.group(1)
    return ""
for dataset in full_data:
    dataset['TitleName'] = dataset['Name'].apply(get_title_name)
print(pd.crosstab(train['TitleName'],train['Sex']))
###### out
Sex        female  male
TitleName              
Capt            0     1
Col             0     2
Countess        1     0
Don             0     1
Dr              1     6
Jonkheer        0     1
Lady            1     0
Major           0     2
Master          0    40
Miss          182     0
Mlle            2     0
Mme             1     0
Mr              0   517
Mrs           125     0
Ms              1     0
Rev             0     6
Sir             0     1
####可以看出不同的titlename中男女还是有区别的。进一步探索titlename对Survived的影响。
####看出上面的离散取值范围还是比较多,所以可以将较少的几类归为一个类别。
train['TitleName'] = train['TitleName'].replace('Mme', 'Mrs')
train['TitleName'] = train['TitleName'].replace('Mlle', 'Miss')
train['TitleName'] = train['TitleName'].replace('Ms', 'Miss')
train['TitleName'] = train['TitleName'].replace(['Lady', 'Countess','Capt', 'Col',\
     'Don', 'Dr', 'Major', 'Rev', 'Sir', 'Jonkheer', 'Dona'], 'Other')
print (train[['TitleName', 'Survived']].groupby('TitleName', as_index=False).mean())
###### out1
  TitleName  Survived
0    Master  0.575000
1      Miss  0.702703
2        Mr  0.156673
3       Mrs  0.793651
4     Other  0.347826

  可以看出TitleName对存活率还是有影响差异的,TitleName总共为了5个类别:Mrs,Miss,Master,Mr,Other。

 

 2.3 特征提取总结

  此赛题是计算每一个属性与响应变量label的影响(存活率)来查看是否选择该属性作为最后模型的输入特征。最后选取出的模型输入特征有Pclass,Sex,CategoricalAge,FamilySize,Parch,SibSp,IsAlone,CategoricalFare,Embarked,TitleName,最后对上述分析进行统一的数据清洗,将train.csv和test.csv统一进行处理,得出新的模型训练样本集。

 

三、XGBoost模型训练

 3.1数据清洗和特征选择

  此步骤主要是根据3中的数据分析来进行编写的。着重点Age的缺失值使用了两种方式进行填充。均值和通过其他清洗的数据特征使用随机森林预测缺失值两种方式。

 

def Passenger_sex(x):
    sex = {'female':0, 'male':1}
    return sex[x]
def Passenger_Embarked(x):
    Embarked = {'S':0, 'C':1, 'Q':2}
    return Embarked[x]
def Passenger_TitleName(x):
    TitleName = {'Mr':0,'Miss':1, 'Mrs':2, 'Master':3, 'Other':4}
    return TitleName[x]

def get_title_name(name):
    title_s = re.search(' ([A-Za-z]+)\.', name)
    if title_s:
        return title_s.group(1)
    return ""

def data_feature_engineering(full_data,age_default_avg=True,one_hot=True):
    """
    :param full_data:全部数据集包括train,test
    :param age_default_avg:age默认填充方式,是否使用平均值进行填充
    :param one_hot: Embarked字符处理是否是one_hot编码还是映射处理
    :return: 处理好的数据集
    """
    for dataset in full_data:
        # Pclass、Parch、SibSp不需要处理

        # sex 0,1
        dataset['Sex'] = dataset['Sex'].map(Passenger_sex).astype(int)

        # FamilySize
        dataset['FamilySize'] = dataset['SibSp'] + dataset['Parch'] + 1

        # IsAlone
        dataset['IsAlone'] = 0
        isAlone_mask = dataset['FamilySize'] == 1
        dataset.loc[isAlone_mask, 'IsAlone'] = 1

        # Fare 离散化处理,6个阶段
        fare_median = dataset['Fare'].median()
        dataset['CategoricalFare'] = dataset['Fare'].fillna(fare_median)
        dataset['CategoricalFare'] = pd.qcut(dataset['CategoricalFare'],6,labels=[0,1,2,3,4,5])

        # Embarked映射处理,one-hot编码,极少部分缺失值处理
        dataset['Embarked'] = dataset['Embarked'].fillna('S')
        dataset['Embarked'] = dataset['Embarked'].astype(str)
        if one_hot:
            # 因为OneHotEncoder只能编码数值型,所以此处使用LabelBinarizer进行独热编码
            Embarked_arr = LabelBinarizer().fit_transform(dataset['Embarked'])
            dataset['Embarked_0'] = Embarked_arr[:, 0]
            dataset['Embarked_1'] = Embarked_arr[:, 1]
            dataset['Embarked_2'] = Embarked_arr[:, 2]
            dataset.drop('Embarked',axis=1,inplace=True)
        else:
            # 字符串映射处理
            dataset['Embarked'] = dataset['Embarked'].map(Passenger_Embarked).astype(int)

        # Name选取称呼Title_name
        dataset['TitleName'] = dataset['Name'].apply(get_title_name)
        dataset['TitleName'] = dataset['TitleName'].replace('Mme', 'Mrs')
        dataset['TitleName'] = dataset['TitleName'].replace('Mlle', 'Miss')
        dataset['TitleName'] = dataset['TitleName'].replace('Ms', 'Miss')
        dataset['TitleName'] = dataset['TitleName'].replace(['Lady', 'Countess', 'Capt', 'Col', \
                                                             'Don', 'Dr', 'Major', 'Rev', 'Sir', 'Jonkheer', 'Dona'],
                                                            'Other')
        dataset['TitleName'] = dataset['TitleName'].map(Passenger_TitleName).astype(int)

        # age —— 缺失值,分段处理
        if age_default_avg:
            # 缺失值使用avg处理
            age_avg = dataset['Age'].mean()
            age_std = dataset['Age'].std()
            age_null_count = dataset['Age'].isnull().sum()
            age_default_list = np.random.randint(low=age_avg - age_std, high=age_avg + age_std, size=age_null_count)

            dataset.loc[np.isnan(dataset['Age']), 'Age'] = age_default_list
            dataset['Age'] = dataset['Age'].astype(int)
        else:
            # 将age作为label,预测缺失的age
            # 特征为 TitleName,Sex,pclass,SibSP,Parch,IsAlone,CategoricalFare,FamileSize,Embarked
            feature_list = ['TitleName', 'Sex', 'Pclass', 'SibSp', 'Parch', 'IsAlone','CategoricalFare',
                            'FamilySize', 'Embarked','Age']
            if one_hot:
                feature_list.append('Embarked_0')
                feature_list.append('Embarked_1')
                feature_list.append('Embarked_2')
                feature_list.remove('Embarked')
            Age_data = dataset.loc[:,feature_list]

            un_Age_mask = np.isnan(Age_data['Age'])
            Age_train = Age_data[~un_Age_mask] #要训练的Age

            # print(Age_train.shape)
            feature_list.remove('Age')
            rf0 = RandomForestRegressor(n_estimators=60,oob_score=True,min_samples_split=10,min_samples_leaf=2,
                                                                   max_depth=7,random_state=10)

            rf0.fit(Age_train[feature_list],Age_train['Age'])

            def set_default_age(age):
                if np.isnan(age['Age']):
                    data_x = np.array(age.loc[feature_list]).reshape(1,-1)
                    age_v = round(rf0.predict(data_x))
                    #age_v = np.round(rf0.predict(data_x))[0]
                    return age_v

                return age['Age']

            dataset['Age'] = dataset.apply(set_default_age, axis=1)

        # pd.cut与pd.qcut的区别,前者是根据取值范围来均匀划分,
        # 后者是根据取值范围的各个取值的频率来换分,划分后的某个区间的频率数相同
        # print(dataset.tail())
        dataset['CategoricalAge'] = pd.cut(dataset['Age'], 5,labels=[0,1,2,3,4])
    return full_data
##特征选择
def data_feature_select(full_data):
    """
    :param full_data:全部数据集
    :return:
    """
    for data_set in full_data:
        drop_list = ['PassengerId','Name','Age','Fare','Ticket','Cabin']
        data_set.drop(drop_list,axis=1,inplace=True)
    train_y = np.array(full_data[0]['Survived'])
    train = full_data[0].drop('Survived',axis=1,inplace=False)
    # print(train.head())
    train_X = np.array(train)
    test_X = np.array(full_data[1])
    return train_X,train_y,test_X        

 

 4.2XGBoost参数介绍

  要熟练的使用XGBoost库一方面需要对XGBoost原理的了解,另一方面需要对XGBoost库的API参数的了解,通用参数

  booster[默认gbtree]:选择每次迭代的模型

   gbtree:基于树的模型 

   gbliner:线性模型 

  nthread:[默认值为最大可能的线程数]

   这个参数用来进行多线程控制,应当输入系统的核数。

   如果你希望使用CPU全部的核,那就不要输入这个参数,算法会自动检测它。

 

  booster:尽管有两种booster可供选择,这里只介绍tree booster,因为它的表现远远胜过linear booster,所以linear booster很少用到

   learning_rate:梯度下降的学习率,一般为0.01~0.2

   min_child_weight:决定最小叶子节点样本权重和,这个参数用于避免过拟合。当它的值较大时,可以避免模型学习到局部的特殊样本;但是如果这个值过高,会导致欠拟合。这个参数需要使用CV来调整

   max_depth:决策树的最大深度,默认为6,这个值也是用来避免过拟合的。max_depth越大,模型会学到更具体更局部的样本;需要使用CV函数来进行调优,典型值:3-10

   max_leaf_nodes:树上最大的叶子数量

   gamma:在节点分裂时,只有分裂后损失函数的值下降了,才会分裂这个节点。Gamma指定了节点分裂所需的最小损失函数下降值。这个参数的值越大,算法越保守。这个参数的值和损失函数息息相关,所以是需要调整的

   subsample和colsample_bytree:随机森林中的两种随机,也是XGBoost中的trick,用于防止过拟合,值为0.5~1,随机采样所占比例,随机列采样比例。

   lambda:L2正则化项,可调参实现。

   scale_pos_weight:在各类别样本十分不平衡时,把这个参数设定为一个正值,可以使算法更快收敛。

 

  学习目标函数

   objective:指定分类回归问题。如binary:logistic

   eval_metric:评价指标

   seeds:随机数种子,调整参数时,随机取同样的样本集

 

import pandas as pd
import numpy as np
import re
from sklearn.preprocessing import LabelBinarizer
from sklearn.ensemble import RandomForestRegressor

import xgboost as xgb
from xgboost.sklearn import XGBClassifier

import matplotlib.pyplot as plt
from sklearn import metrics
from sklearn.model_selection import GridSearchCV

train_file = "C:\\Users\\Administrator\\Desktop\\python\\data\\Titanic\\train.csv"
test_file = "C:\\Users\\Administrator\\Desktop\\python\\data\\Titanic\\test.csv"
test_result_file = "C:\\Users\\Administrator\\Desktop\\python\\data\\Titanic\\gender_submission.csv"
train = pd.read_csv(train_file)
test = pd.read_csv(test_file)
test_y = pd.read_csv(test_result_file)

def Passenger_sex(x):
    sex = {'female':0, 'male':1}
    return sex[x]

def Passenger_Embarked(x):
    Embarked = {'S':0, 'C':1, 'Q':2}
    return Embarked[x]

def get_title_name(name):
    title_s = re.search(' ([A-Za-z]+)\.', name)
    if title_s:
        return title_s.group(1)
    return ""

def Passenger_TitleName(x):
    TitleName = {'Mr':0,'Miss':1, 'Mrs':2, 'Master':3, 'Other':4}
    return TitleName[x]

#https://github.com/JianWenJun/MLDemo/blob/master/ML/DecisionTree/xgboost_demo.py
 
#数据清洗和特征选择
##此步骤主要是根据数据分析来进行编写的。着重点Age的缺失值使用了两种方式进行填充。均值和通过其他清洗的数据特征使用随机森林预测缺失值两种方式。
def data_feature_engineering(full_data, age_default_avg=True, one_hot=True):
    """
    :param full_data:全部数据集,包括train,test
    :param age_default_avg: age默认填充方式,是否使用平均值进行填充
    :param one_hot: Embarked字符处理是否是one_hot编码还是映射处理
    :return:处理好的数据集
    """
    for dataset in full_data:
        #Pclass,Parch,SibSp不需要处理
        
        #Sex没有空值,用 0,1做映射,{'female':0, 'male':1}
        dataset['Sex'] = dataset['Sex'].map(Passenger_sex).astype(int)
        
        #SibSp和Parch均没有空值,组合特征FamilySize
        dataset['FamilySize'] = dataset['SibSp'] + dataset['Parch'] + 1
        
        #根据FamilySize大小增加新特征IsAlone
        dataset['IsAlone'] = 0
        isAlone_mask = dataset['FamilySize'] == 1
        dataset.loc[isAlone_mask,'IsAlone'] = 1
        
        #Fare的test数据存在一个空值,先填充,然后Fare离散化处理,6个阶段
        fare_median = dataset['Fare'].median()
        dataset['CategoricalFare'] = dataset['Fare'].fillna(fare_median)
        dataset['CategoricalFare'] = pd.qcut(dataset['CategoricalFare'],6,labels=[0,1,2,3,4,5])
        
        #Embarked存在空值,众数填充,映射处理或者one-hot 编码
        dataset['Embarked'] = dataset['Embarked'].fillna('S')
        dataset['Embarked'] = dataset['Embarked'].astype(str)
        if one_hot:
            #因为OneHotEncoder只能编码数值型,所以此处使用LabelBinarizer进行独热编码
            Embarked_arr = LabelBinarizer().fit_transform(dataset['Embarked'])
            dataset['Embarked_0'] = Embarked_arr[:, 0]
            dataset['Embarked_1'] = Embarked_arr[:, 1]
            dataset['Embarked_2'] = Embarked_arr[:, 2]
            dataset.drop('Embarked', axis=1, inplace=True)
        else:
            #字符映射处理
            dataset['Embarked'] = dataset['Embarked'].map(Passenger_Embarked).astype(int)
            
        #Name选取称呼Title_name,{'Mr':0,'Miss':1, 'Mrs':2, 'Master':3, 'Other':4}
        dataset['TitleName'] = dataset['Name'].apply(get_title_name)
        dataset['TitleName'] = dataset['TitleName'].replace('Mme','Mrs')
        dataset['TitleName'] = dataset['TitleName'].replace('Mlle', 'Miss')
        dataset['TitleName'] = dataset['TitleName'].replace('Ms', 'Miss')
        dataset['TitleName'] = dataset['TitleName'].replace(['Lady', 'Countess', 'Capt', 'Col', \
                                                             'Don', 'Dr', 'Major', 'Rev', 'Sir', 'Jonkheer', 'Dona'],
                                                            'Other')
        dataset['TitleName'] = dataset['TitleName'].map(Passenger_TitleName).astype(int)
        
        #age 缺失值,可以使用随机值填充,也可以用RF预测
        if age_default_avg:
            #缺失值使用avg处理
            age_avg = dataset['Age'].mean()
            age_std = dataset['Age'].std()
            age_null_count = dataset['Age'].isnull().sum()
            age_default_list = np.random.randint(low=age_avg-age_std, high=age_avg+age_std, size=age_null_count)
            dataset.loc[np.isnan(dataset['Age']),'Age'] = age_default_list
            dataset['Age'] = dataset['Age'].astype(int)
            
        else:
            #将age作为label,使用RF的age
            #特征为TitleName,Sex,pclass,SibSP,Parch,IsAlone,CategoricalFare,FamileSize,Embarked
            feature_list = ['TitleName', 'Sex', 'Pclass', 'SibSp', 'Parch', 'IsAlone', 'CategoricalFare',
                            'FamilySize', 'Embarked', 'Age']
            if one_hot:
                feature_list.append('Embarked_0')
                feature_list.append('Embarked_1')
                feature_list.append('Embarked_2')
                feature_list.remove('Embarked')
                
            Age_data = dataset.loc[:, feature_list]
            un_Age_mask = np.isnan(Age_data['Age'])
            Age_train = Age_data[~un_Age_mask]#要训练的Age

            feature_list.remove('Age')
            
            rf0 = RandomForestRegressor(n_estimators=60, oob_score=True, min_samples_split=10, min_samples_leaf=2, max_depth=7, random_state=10)
            rf0.fit(Age_train[feature_list], Age_train['Age'])
            
            #这里处理的很巧,注意一下
            def set_default_age(age):
                if np.isnan(age['Age']):
                    data_x = np.array(age.loc[feature_list]).reshape(1,-1)
                    #age_v = round(rf0.predict(data_x))
                    age_v = np.round(rf0.predict(data_x))[0]
                    print(age_v)
                    return age_v
                return age['Age']
 
            dataset['Age'] = dataset.apply(set_default_age, axis=1)
             
        # pd.cut与pd.qcut的区别,前者是根据取值范围来均匀划分
        # 后者是根据取值范围的各个取值的频率来换分,划分后的某个区间的频率数相同
        dataset['CategoricalAge'] = pd.cut(dataset['Age'], 5, labels=[0,1,2,3,4])
    return full_data


##特征选择
def data_feature_select(full_data):
    """
    :param full_data:全部数据集
    :return:
    """
    for data_set in full_data:
        drop_list = ['PassengerId', 'Name', 'Age', 'Fare', 'Ticket', 'Cabin']
        data_set.drop(drop_list, axis=1, inplace=True)
        
    train_y = np.array(full_data[0]['Survived'])
    train = full_data[0].drop('Survived', axis=1, inplace=False)
     
    train_X = np.array(train)
    test_X = np.array(full_data[1])
    return train_X, train_y, test_X

full_data = [train,test]
full_data = data_feature_engineering(full_data, age_default_avg=True, one_hot=False)
train_X, train_y, test_X = data_feature_select(full_data)

def modelfit(alg,dtrain_x,dtrain_y,useTrainCV=True,cv_flods=5,early_stopping_rounds=50):
    """
    :param alg: 初始模型
    :param dtrain_x:训练数据X
    :param dtrain_y:训练数据y(label)
    :param useTrainCV: 是否使用cv函数来确定最佳n_estimators
    :param cv_flods:交叉验证的cv数
    :param early_stopping_rounds:在该数迭代次数之前,eval_metric都没有提升的话则停止
    """
    if useTrainCV:
        xgb_param = alg.get_xgb_params()
        xgtrain = xgb.DMatrix(dtrain_x,dtrain_y)
        print(alg.get_params()['n_estimators'])
        cv_result = xgb.cv(xgb_param,xgtrain,num_boost_round = alg.get_params()['n_estimators'],
                           nfold = cv_flods, metrics = 'auc', early_stopping_rounds=early_stopping_rounds)

        print(cv_result)
        print(cv_result.shape[0])
        alg.set_params(n_estimators=cv_result.shape[0])
 
    # train data
    alg.fit(train_X,train_y,eval_metric='auc')
 
    #predict train data
    train_y_pre = alg.predict(train_X)
 
    print ("\nModel Report")
    print ("Accuracy : %.4g" % metrics.accuracy_score(train_y, train_y_pre))
 
    feat_imp = pd.Series(alg.get_booster().get_fscore()).sort_values(ascending=False)
    feat_imp.plot(kind = 'bar',title='Feature Importance')
    plt.ylabel('Feature Importance Score')
    plt.show()
    print(alg)
    
def xgboost_change_param(train_X, train_y):
    ######Xgboost 调参########
     
    #step1 确定学习速率和迭代次数n_estimators
    xgb1 = XGBClassifier(learning_rate=0.1,
                        booster='gbtree',
                         n_estimators=300,
                         max_depth=4,
                         min_child_weight=1,
                         gamma=0,
                         subsample=0.8,
                         colsample_bytree=0.8,
                         objective='binary:logistic',
                         nthread=2,
                         scale_pos_weight=1,
                         seed=10)
    #最佳n_estimators = 11,earning_rate=0.1
    modelfit(xgb1, train_X, train_y, early_stopping_rounds=45)
    
    #step2调试参数min_child_weight以及max_depth
    param_test1 = {'max_depth':range(3,8,1),
                    'min_child_weight':range(1,6,2)}
    
    gsearch1 = GridSearchCV(estimator=XGBClassifier(learning_rate=0.1,
                                                    n_estimators=11,
                                                    max_depth=4,
                                                    min_child_weight=1,
                                                    gamma=0,
                                                    subsample=0.8,
                                                    colsample_bytree=0.8,
                                                    objective='binary:logistic',
                                                    nthreads=2,
                                                    scale_pos_weight=1,seed=10),
                            param_grid=param_test1,scoring='roc_auc',n_jobs=1,cv=5)
    
    gsearch1.fit(train_X,train_y)
    
    print(gsearch1.best_params_, gsearch1.best_score_)
    #最佳max_depth = 5 min_child_weight=1
    
    #step3 gamma参数调优
    param_test2 = {'gamma': [i/10.0 for i in range(0,5)]}
    gsearch2 = GridSearchCV(estimator=XGBClassifier(learning_rate=0.1,n_estimators=11,
                                                    max_depth=5,min_child_weight=1,gamma=0,
                                                    subsample=0.8,colsample_bytree=0.8,
                                                    objective='binary:logistic',nthread=2,
                                                    scale_pos_weight=1,seed=10),
                            param_grid=param_test2,
                            scoring='roc_auc',
                            cv=5)
    gsearch2.fit(train_X, train_y)
    print(gsearch2.best_params_, gsearch2.best_score_)
    #最佳 gamma = 0.2

    #step4 调整subsample 和 colsample_bytrees参数
    param_test3 = {'subsample': [i/10.0 for i in range(6,10)],
                   'colsample_bytree': [i/10.0 for i in range(6,10)]}
    gsearch3 = GridSearchCV(estimator=XGBClassifier(learning_rate=0.1,n_estimators=11,
                                                    max_depth=5,min_child_weight=1,gamma=0.2,
                                                    subsample=0.8,colsample_bytree=0.8,
                                                    objective='binary:logistic',nthread=2,
                                                    scale_pos_weight=1,seed=10),
                            param_grid=param_test3,
                            scoring='roc_auc',
                            cv=5
                            )
    gsearch3.fit(train_X, train_y)
    print(gsearch3.best_params_, gsearch3.best_score_)
    #最佳'subsample': 0.7, 'colsample_bytree': 0.8
    
    #step5 正则化参数调优
    param_test4= { 'reg_alpha':[1e-5, 1e-2, 0.1, 1, 100] }
    gsearch4= GridSearchCV(estimator=XGBClassifier(learning_rate=0.1,n_estimators=11,
                                                    max_depth=5,min_child_weight=1,gamma=0.2,
                                                    subsample=0.7,colsample_bytree=0.8,
                                                    objective='binary:logistic',nthread=2,
                                                    scale_pos_weight=1,seed=10),
                            param_grid=param_test4,
                            scoring='roc_auc',
                            cv=5
                            )
    gsearch4.fit(train_X, train_y)
    print(gsearch4.grid_scores_, gsearch4.best_params_, gsearch4.best_score_)
    #reg_alpha:0.01
    
    param_test5 ={ 'reg_lambda':[1e-5, 1e-2, 0.1, 1, 100] }
    gsearch5= GridSearchCV(estimator=XGBClassifier(learning_rate=0.1,n_estimators=11,
                                                    max_depth=5,min_child_weight=1,gamma=0.2,
                                                    subsample=0.7,colsample_bytree=0.8,
                                                    objective='binary:logistic',nthread=2,reg_alpha=0.01,
                                                    scale_pos_weight=1,seed=10),
                            param_grid=param_test5,
                            scoring='roc_auc',
                            cv=5
                            )
    gsearch5.fit(train_X, train_y)
    print(gsearch5.grid_scores_, gsearch5.best_params_, gsearch5.best_score_)
    #reg_lambda:0.1
    
    
# XGBoost调参
xgboost_change_param(train_X, train_y)

#parameters at last
xgb1 = XGBClassifier(learning_rate=0.1,n_estimators=11,
                    max_depth=5,min_child_weight=1,
                    gamma=0.2,subsample=0.7,
                    colsample_bytree=0.8, objective='binary:logistic',reg_alpha=0.01,reg_lambda=0.1,
                    nthread=2, scale_pos_weight=1,seed=10)
xgb1.fit(train_X,train_y)
 
y_test_pre = xgb1.predict(test_X)
y_test_true = np.array(test_y['Survived'])
print ("the xgboost model Accuracy : %.4g" % metrics.accuracy_score(y_pred=y_test_pre, y_true=y_test_true))
'''
{'max_depth': 5, 'min_child_weight': 1} 0.8681463854050605
{'gamma': 0.2} 0.8691675488262394
{'colsample_bytree': 0.8, 'subsample': 0.7} 0.8716601827161007
[mean: 0.87166, std: 0.02694, params: {'reg_alpha': 1e-05}, mean: 0.87198, std: 0.02694, params: {'reg_alpha': 0.01}, mean: 0.87039, std: 0.02804, params: {'reg_alpha': 0.1}, mean: 0.87094, std: 0.02296, params: {'reg_alpha': 1}, mean: 0.77439, std: 0.02898, params: {'reg_alpha': 100}] {'reg_alpha': 0.01} 0.8719802934262043
[mean: 0.87162, std: 0.02466, params: {'reg_lambda': 1e-05}, mean: 0.87169, std: 0.02504, params: {'reg_lambda': 0.01}, mean: 0.87218, std: 0.02538, params: {'reg_lambda': 0.1}, mean: 0.87198, std: 0.02694, params: {'reg_lambda': 1}, mean: 0.84922, std: 0.01254, params: {'reg_lambda': 100}] {'reg_lambda': 0.1} 0.8721773740150154
the xgboost model Accuracy : 0.9282
'''

 

Reference:

  https://blog.csdn.net/u014732537/article/details/80055227

  https://github.com/JianWenJun/MLDemo/blob/master/ML/DecisionTree/xgboost_demo.py

posted @ 2018-11-08 16:26  1直在路上1  阅读(1504)  评论(0编辑  收藏  举报