Udemy: Final Project Review

1、数据初始化，以及删除含缺失值行

import pandas as pd
#不读取首列为列名；将?替换为NaN
masses=pd.read_csv('mammographic_masses.data.txt',header=None,na_values='?')
masses.head()


masses.columns=["BI_RADS","age","shape","margin","density","severity"]
masses.head()


#做描述性统计
masses.describe()


#通过展示含缺失值行，判断缺失值在变量分布是否均匀，决定是否删除含缺失值行，避免因某一特征缺失值较多造成该特征的信息损失
masses.loc[(masses['age'].isnull())|
           (masses['shape'].isnull())|
           (masses['margin'].isnull())|
           (masses['density'].isnull())]


#发现缺失值分布较均匀，删除含缺失值行
#masses=masses.dropna(axis=0)
#善用inplace选项，axis=1表示删除含缺失值列
masses.dropna(axis=0, inplace=True)
masses.describe()


#使用values直接将dataframe转化为numpy
#我的做法:dataframe.to_numpy
N_masses=masses.to_numpy()
Feature=N_masses[:,0:4]
Severity=N_masses[:,5]
#values做法
all_features = masses.iloc[:,0:4].values
all_classes = masses.iloc[:,5].values
feature_names = list(masses.columns.values)


#Z_score标准化：将数据标准化为标准正态分布
from sklearn import preprocessing
scaler=preprocessing.StandardScaler()
all_features_scaled = scaler.fit_transform(all_features)
all_features_scaled

2、Decision Trees

(1) Train_test_split分割样本

#train_test_split分割样本


#我的方法手动分割样本
#train=N_masses[:620,:]
#test=N_masses[620:,:]

#教程方法：自动分割样本, train_test_split
from sklearn.model_selection import cross_val_score, train_test_split
import numpy
numpy.random.seed(1234)

(training_inputs,
 testing_inputs,
 training_classes,
 testing_classes) = train_test_split(all_features, all_classes, train_size=0.75, random_state=1)
#random_state=1, 表示多次split的train和test一致

(2) Train the decision tree

#训练模型
from sklearn import tree

clf = tree.DecisionTreeClassifier(random_state=1)
clf = clf.fit(training_inputs,training_classes)

#显示Decision Tree
from IPython.display import Image
from sklearn.externals.six import StringIO
import pydotplus
dot_data = StringIO()
tree.export_graphviz(clf, out_file=dot_data, 
                    feature_names=feature_names)
graph = pydotplus.graph_from_dot_data(dot_data.getvalue())
Image(graph.create_png())

(3) Measuring the result

clf.score(testing_inputs, testing_classes)

3、K-Fold cross validation + Decision Tree

(1) 简介：

        模型参数选择往往取决于test的结果，即造成了前视偏差和overfitting。因此，我们使用validation选择最佳参数，并将该参数用于最终的test。

        但是，因为将dataset分为三部分：train、validation以及test，造成可用样本量减少，并且train和validation样本量的随机选择也会影响到test的结果。

        因此，K-fold cross validation应运而生。通过将trainK等分，并从中仍迭代选择一等分作为validation，根据其平均值选择模型参数，既避免了前视偏差，也解决了样本量不足。

(2) 代码

#K-Fold方法直接给定整个数据集，且使用了标准化的数据

from sklearn.model_selection import cross_val_score, train_test_split

clf = DecisionTreeClassifier(random_state=1)

cv_scores= cross_val_score(clf, all_features_scaled, all_classes, cv=10)

#计算十种validation抽样方法得分均值
cv_scores.mean()

 4、K-Fold cross validation + RandomForest

(1) 简介

        随机森林就是通过集成学习的思想将多棵树集成的一种算法，它的基本单元是决策树，而它的本质属于机器学习的一大分支–集成学习（Ensemble Learning）方法。

  从直观角度来解释，每棵树都是一个分类器（假设现在为分类问题），那么对于一个输入样本，N棵树会有N个分类结果。而随机森林集成了所有的分类投票结果，将投票次数最多的类别指定为最终的输出，这就是一种最简单的Bagging思想。

(2)  随机森林的生成过程

从原始训练集中使用Bootstraping方法随机有放回采用选出m个样本，共进行n_tree次采样，生成n_tree个训练集；

对于n_tree个训练集，分别训练n_tree个决策树模型；

对于单个决策树模型，假设训练样本特征的个数为n，那么每次分裂时根据信息增益/信息增益比/Gini指数选择最好的特征进行分裂；

每棵树都一直这样分裂下去，直到该节点的所有训练样例都属于同一类。在决策树的分裂过程中不需要剪枝；

将生成的多棵决策树组成随机森林。对于分类问题，按多棵树分类器投票决定最终分类结果；对于回归问题，由多棵树预测的均值决定最终预测结果。

(3) 代码

from sklearn.ensemble import RandomForestClassifier

clf = RandomForestClassifier(n_estimators=30, random_state=1)

cv_scores= cross_val_score(clf, all_features_scaled, all_classes, cv=10)

cv_scores.mean()

5、SVM.SVC: Support Vector Classification

from sklearn import svm

svc = svm.SVC(kernel='linear', C=1)#分为两类，设置C=1
cv_scores= cross_val_score(svc, all_features_scaled, all_classes, cv=10)
cv_scores.mean()

svc = svm.SVC(kernel='rbf', C=1)
cv_scores= cross_val_score(svc, all_features_scaled, all_classes, cv=10)
cv_scores.mean()

svc = svm.SVC(kernel='sigmoid', C=1)
cv_scores= cross_val_score(svc, all_features_scaled, all_classes, cv=10)
cv_scores.mean()

svc = svm.SVC(kernel='poly', C=1)#分为两类，设置C=1
cv_scores= cross_val_score(svc, all_features_scaled, all_classes, cv=10)
cv_scores.mean()

6、K-Nearest-Neighbors

from sklearn.neighbors import KNeighborsClassifier

neigh = KNeighborsClassifier(n_neighbors=10)

cv_scores= cross_val_score(neigh, all_features_scaled, all_classes, cv=10)

cv_scores.mean()

#select best K-value
for K in range (1,51):
    neigh = KNeighborsClassifier(n_neighbors=K)
    cv_scores= cross_val_score(neigh, all_features_scaled, all_classes, cv=10)
    print (K, cv_scores.mean())

7、Naive_Bayes.MultinomialNB: 先验为多项式分布的朴素贝叶斯

(1) 算法介绍

https://blog.csdn.net/TeFuirnever/article/details/100108341

(2) 代码

from sklearn.preprocessing import MinMaxScaler
from sklearn.naive_bayes import MultinomialNB

#Naive_Bayes需归一化处理
scaler = MinMaxScaler() 
all_features_minmax = scaler.fit_transform(all_features)

cla = MultinomialNB()
cv_scores= cross_val_score(cla, all_features_minmax, all_classes, cv=10)
cv_scores.mean()

8、Logistic Regression

(1) 算法

https://blog.csdn.net/guoziqing506/article/details/81328402

(2) 代码

from sklearn.linear_model import LogisticRegression

clf = LogisticRegression()

cv_scores= cross_val_score(clf, all_features_scaled, all_classes, cv=10)

cv_scores.mean()

9、Neural Network

from tensorflow.keras.wrappers.scikit_learn import KerasClassifier
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout

def create_model():
    model = Sequential()
    model.add(Dense(10, input_dim=4, kernel_initializer='normal', activation='relu'))
    model.add(Dense(1, kernel_initializer='normal',dropout=0.8, activation='softmax'))
    model.compile(loss='binary_crossentropy',optimizer=RMSprop(),metrics=['accuracy'])
    return model

estimator = KerasClassifier(build_fn=create_model, epochs=100, verbose=0)
cv_scores= cross_val_score(clf, all_features_scaled, all_classes, cv=10)
cv_scores.mean()