Kaggle Titanic补充篇

1.关于年龄Age

除了利用平均数来填充，还可以利用正态分布得到一些随机数来填充，首先得到已知年龄的平均数mean和方差std,然后生成[ mean-std,  mean+std ]之间的随机数，然后利用这些随机值填充缺失的年龄。

2.关于票价Fare

预处理：训练集不缺，测试集缺失1个，用最高频率值填充

Fare_freq = test.Fare.dropna().mode()[0]  # 找出非缺失值中的所有最高频值，取第一个
for dataset in train_test_data:
    dataset['Fare'] = dataset['Fare'].fillna(Fare_freq)

特征工程：由于Fare分布非常不均，所以这里不用cut函数，而是qcut，因为它可以根据样本分位数对数据进行面元划分，可以使得到的类别中数目基本一样。

train['FareBins'] = pd.qcut(train['Fare'], 5)  # 按照分位数切成5份
train[['FareBins', 'Survived']].groupby(['FareBins'], as_index = False).mean().sort_values(by = 'Survived', ascending = True)

fare = data_train.groupby(['FareBins'])
  fare_l = (fare.sum()/fare.count())['Survived']

fare_l.plot(kind='bar')

     

上图左为cut函数结果，中为qcut函数结果，右图为各年龄段的生存率，升序排列，可以看到生存率基本上按照票价增加而增加。然后在数据集中新增FareBins取值0~4，然后删除Fare项。

3.关于亲人SibSp和Parch

新增Family特征为SibSp和Parch的和，删掉SibSp和Parch

dataset["FamilySize"] = dataset['SibSp'] + dataset['Parch']

train[['FamilySize', 'Survived']].groupby(['FamilySize'], as_index = False).mean().sort_values(by = 'Survived', ascending = False)

此图是亲人数目与生存率关系，表明有1~3个亲属生存率更高些。所以我考虑将1~3单独作为一类。

4.关于姓名

取出姓名中的’属性‘，例如有：Mr、Miss、Dr、Major等

df =data_train.Name.apply(lambda x: x.split(',')[1].split('.')[0].strip()) 

df.value_counts().plot(kind='bar')

train_test_data = [train, test]  # 将训练集和测试集合并处理
for dset in train_test_data:   # 对于人数较少的属性将其统一命名为‘Lamped’
    dset["Title"] = dset["Title"].replace(["Melkebeke", "Countess", "Capt", "the Countess", "Col", "Don",
                                         "Dr", "Major", "Rev", "Sir", "Jonkheer", "Dona"] , "Lamped")
    dset["Title"] = dset["Title"].replace(["Lady", "Mlle", "Ms", "Mme"] , "Miss")  # 将女性称号合并

train[['Title', 'Survived']].groupby(['Title'], as_index = False).mean().sort_values(by = 'Survived', ascending = False)

然后将这类特征作为新特征取值0~4加入，删掉原有特征Name。

按照以上思路跑了另一个例子，结果并没有提高多少

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from pandas import Series, DataFrame
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import classification_report
from learning_curve import *
from pylab import mpl
from sklearn.model_selection import cross_val_score
mpl.rcParams['font.sans-serif'] = ['SimHei']  #使得plt操作可以显示中文
from sklearn.feature_extraction import DictVectorizer

data_train = pd.read_csv('train.csv')
data_test = pd.read_csv('test.csv')
feature = ['Pclass','Age','Sex','Fare','Embarked','SibSp','Parch','Cabin','Name']

X_train = data_train[feature]
y_train = data_train['Survived']

X_test = data_test[feature]

X_train_test = [X_train, X_test]

"""
填充操作
"""
# 填充Age, Cabin:训练集合测试集都缺，且填补方式一样
def age_fill(X):
    for df in X:
        age_mean = df['Age'].dropna().mean()
        age_std = df['Age'].dropna().std()
        age = np.random.randint(age_mean-age_std, age_mean+age_std, size=df['Age'].isnull().sum())
        df['Age'][np.isnan(df['Age'])] = age
        df['Cabin'][df['Cabin'].isnull()] = 0
        df['Cabin'][df['Cabin']!=0] = 1
    return X

age_fill(X_train_test)

# 填充Embarked:只有训练集缺一个
X_train['Embarked'][X_train['Embarked'].isnull()] = X_train.Embarked.dropna().mode()[0]

# 填充Fare:只有测试集缺一个
X_test['Fare'][X_test['Fare'].isnull()] = X_test.Fare.dropna().mode()[0]

"""
特征工程
"""
# Pclass
def Pclass_f(df):
    dummies_Pclass = pd.get_dummies(df['Pclass'], prefix='Pclass')
    df = pd.concat([df, dummies_Pclass], axis=1)
    df.drop(['Pclass'], axis=1, inplace=True)
    return df

X_train = Pclass_f(X_train)
X_test = Pclass_f(X_test)

# Sex
def Sex_f(df):
    dummies_Sex = pd.get_dummies(df['Sex'], prefix='Sex')
    df = pd.concat([df, dummies_Sex], axis=1)
    df.drop(['Sex'], axis=1, inplace=True)
    return df

X_train = Sex_f(X_train)
X_test = Sex_f(X_test)

# Embarked
def Embarked_f(df):
    dummies_Embarked = pd.get_dummies(df['Embarked'], prefix='Embarked')
    df = pd.concat([df, dummies_Embarked], axis=1)
    df.drop(['Embarked'], axis=1, inplace=True)
    return df

X_train = Embarked_f(X_train)
X_test = Embarked_f(X_test)

X_train_test = [X_train, X_test]
# SibSp+Parch
for df in X_train_test:
    df['Family'] = df['SibSp'] + df['Parch']
    df.drop(['SibSp', 'Parch'], axis=1, inplace=True)

# Age
for df in X_train_test:
    df['Age_new'] = (df['Age'] / 8).astype(int)
    df.drop(['Age'], axis=1, inplace=True)

# Fare
for df in X_train_test:
    df['Fare_new'] = pd.qcut(df['Fare'], 5)
    df.loc[df['Fare'] <= 7.854, 'Fare'] = 0
    df.loc[(df['Fare'] > 7.84) & (df['Fare'] <= 10.5), 'Fare'] = 1
    df.loc[(df['Fare'] > 10.5) & (df['Fare'] <= 21.679), 'Fare'] = 2
    df.loc[(df['Fare'] > 21.679) & (df['Fare'] <= 39.688), 'Fare'] = 3
    df.loc[(df['Fare'] > 39.688) & (df['Fare'] <= 5512.329), 'Fare'] = 4
    df.drop(['Fare_new'], axis=1, inplace=True)

# Name
for df in X_train_test:
    df['Name_new'] = df['Name'].apply(lambda x: x.split(',')[1].split('.')[0].strip())
    df["Name_new"] = df["Name_new"].replace(["Melkebeke", "Countess", "Capt", "the Countess", "Col", "Don",
                                               "Dr", "Major", "Rev", "Sir", "Jonkheer", "Dona"], "Lamped")
    df["Name_new"] = df["Name_new"].replace(["Lady", "Mlle", "Ms", "Mme"], "Miss")
    df['Name_new'] = df['Name_new'].map({'Mr': 1, 'Miss': 2, 'Mrs': 3, 'Master': 4, 'Lamped': 5}).astype(int)
    df.drop(['Name'], axis=1, inplace=True)

"""
MACHINE LEARNING
"""
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score
from sklearn.linear_model import LogisticRegression, Perceptron, SGDClassifier
from sklearn.svm import SVC, LinearSVC
from sklearn.ensemble import RandomForestClassifier, AdaBoostClassifier, BaggingClassifier
from sklearn.tree import DecisionTreeClassifier
from sklearn.neighbors import KNeighborsClassifier
from sklearn.naive_bayes import GaussianNB

dec = LogisticRegression(C=1.2, penalty='l2', tol=1e-6, max_iter=150)
# dec = GaussianNB()
# dec = DecisionTreeClassifier()
# dec = RandomForestRegressor(n_estimators = 100)
# dec =DecisionTreeClassifier()
# dec = RandomForestClassifier(n_estimators=100)
# dec = SVC(C=1.3,kernel='poly')
# dec = LinearSVC()
# dec= SGDClassifier(max_iter=8, tol=None)

dec.fit(X_train, y_train)
y_pre = dec.predict(X_test)

# 交叉验证
X_cro = X_train.as_matrix()
y_cro = y_train.as_matrix()
# est = LogisticRegression(C=1.0, penalty='l1', tol=1e-6)
scores = cross_val_score(dec, X_cro, y_cro, cv=5)
print(scores)
print(scores.mean())

"""
KNN
"""
# k_range = list(range(1, 30))
# scores_knn = []
# for k in k_range:
#     knn = KNeighborsClassifier(n_neighbors=k)
#     knn.fit(X_train, y_train)
#     scores_knn.append(cross_val_score(knn, X_cro, y_cro, cv=5).mean())
#     print(k, scores_knn[k-1])
#
# plt.plot(k_range, scores_knn)
# plt.xlabel('K Values')
# plt.ylabel('Accuracy')

"""
Perceptron
"""
# scores_P = []
# for i in range(1,30):
#     clf = Perceptron(max_iter=i, tol=None)
#     clf.fit(X_train, y_train)
#     scores_P.append(cross_val_score(clf, X_cro, y_cro, cv=5).mean())
#     print(i, scores_P[i-1])
#
# # Plot
# plt.plot(range(1,30), scores_P)
# plt.xlabel('max_iter')
# plt.ylabel('Accuracy')

"""
AdaBoost
"""
# e_range = list(range(1, 25))
# scores_A = []
# for est in e_range:
#     ada = AdaBoostClassifier(n_estimators=est)
#     ada.fit(X_train, y_train)
#     scores_A.append(cross_val_score(ada, X_cro, y_cro, cv=5).mean())
#     print(i, scores_A[est-1])
#
# plt.plot(e_range, scores_A)
# plt.xlabel('estimator values')
# plt.ylabel('Accuracy')

"""
Bagging
"""
# e_range = list(range(1, 30))
# scores_B = []
# for est in e_range:
#     ada = BaggingClassifier(n_estimators=est)
#     ada.fit(X_train, y_train)
#     scores_B.append(cross_val_score(ada, X_cro, y_cro, cv=5).mean())
#     print(i, scores_B[est-1])
#
# plt.plot(e_range, scores_B)
# plt.xlabel('estimator values')
# plt.ylabel('Accuracy')

# 学习曲线
plot_learning_curve(dec, u"学习曲线", X_train, y_train)

# 查看各个特征的相关性
columns = list(X_train.columns)
plt.figure(figsize=(8,8))
plot_df = pd.DataFrame(dec.coef_.ravel(), index=columns)
plot_df.plot(kind='bar')
plt.show()

# 保存结果
# result = pd.DataFrame({'PassengerId':data_test['PassengerId'].as_matrix(), 'Survived':y_pre.astype(np.int32)})
# result.to_csv("my_logisticregression_1.csv", index=False)

# Logistic Regression  [ 0.82681564  0.82122905  0.78089888  0.82022472  0.83050847]  0.815935352
# Guassian Naive Bayes [ 0.73184358  0.73184358  0.75842697  0.79775281  0.80225989]   0.764425362625
# Decision Trees  [ 0.77094972  0.74860335  0.80898876  0.7752809   0.84180791] 0.78912612903
# Random Forest  [ 0.81564246  0.77653631  0.83146067  0.80898876  0.84745763] 0.816017167254
# SVM [ 0.84916201  0.81564246  0.81460674  0.80337079  0.85875706] 0.828307811902
# Linear SVC  [ 0.81564246  0.82681564  0.79213483  0.81460674  0.84180791] 0.81820151663933
# Stochastic Gradient Decent (SGD) [ 0.7877095   0.79329609  0.7247191   0.79213483  0.7740113 ] 0.774374163722
# KNN          MAX:  0.82722180788832633
# Perceptron   MAX:  0.79691590180393057
# AdaBoost     MAX:  0.8260603371814963
# Bagging      MAX:  0.81936277456417805

# dec = SVC(C=1.2,kernel='poly')  [ 0.84916201  0.82681564  0.82022472  0.79775281  0.8700565 ] 0.832802335779