机器学习——信用卡反欺诈案例

 1 import numpy as np
 2 import pandas as pd
 3 from pandas import Series, DataFrame
 4 import matplotlib.pyplot as plt
 5 from sklearn.preprocessing import StandardScaler
 6 from imblearn.over_sampling import SMOTE
 7 from sklearn.ensemble import GradientBoostingClassifier
 8 from sklearn.model_selection import train_test_split
 9 from sklearn.linear_model import LogisticRegression
10 from sklearn.metrics import confusion_matrix
11 import itertools
12 from sklearn.model_selection import GridSearchCV
13 from sklearn.metrics import auc, roc_curve

 1 def plot_confusion_matrix(cm, classes,
 2                           title='Confusion matrix',
 3                           cmap=plt.cm.Blues):
 4     """
 5     This function prints and plots the confusion matrix.
 6     """
 7     plt.imshow(cm, interpolation='nearest', cmap=cmap)
 8     plt.title(title)
 9     plt.colorbar()
10     tick_marks = np.arange(len(classes))
11     plt.xticks(tick_marks, classes, rotation=0)
12     plt.yticks(tick_marks, classes)
13 
14     threshold = cm.max() / 2.
15     for i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])):
16         plt.text(j, i, cm[i, j],
17                  horizontalalignment="center",
18                  color="white" if cm[i, j] > threshold else "black")  # 若对应格子上面的数量不超过阈值则，上面的字体为白色，为了方便查看
19 
20     plt.tight_layout()
21     plt.ylabel('True label')
22     plt.xlabel('Predicted label')
23     plt.show()

1 credit = pd.read_csv('./creditcard.csv')
2 
3 print('原始行列 >>>>', credit.shape)  # (284807行, 31列)
4 # print(credit.head())  # 前5行
5 # print(credit.dtypes)  # 查看特征(列)类型。结果：数据类型只有float64和int64
6 # print(credit.isnull().any())    # 判断是否有缺失值。结果：无缺失值，方便后续处理
7 # print(credit.info())  # 查看数据集详细信息(类型，占用大小，缺失值，行列等)

 1 # c_counts = credit['Class'].value_counts()
 2 # print(c_counts, type(c_counts))  # 对Class列分类统计,并判断类型
 3 # print(c_counts.index, c_counts.values)  # 提取索引和值
 4 '''
 5 结果：
 6 0    284315
 7 1       492
 8 Name: Class, dtype: int64
 9 Name: Class, dtype: int64 <class 'pandas.core.series.Series'>
10 Int64Index([0, 1], dtype='int64') [284315    492]
11 '''
12 
13 # 对c_counts作图进行分析
14 # plt.figure(figsize=(10, 6))
15 # 饼图：两种作图方式
16 # ax = plt.subplot(121)
17 # c_counts是pandas的Series类型，pandas可以使用plot快速作图
18 # c_counts.plot(kind='pie', autopct='%0.3f%%', ax=ax)
19 # plt.pie(c_counts, autopct='%0.3f%%')
20 
21 # 柱状图：两种作图方式
22 # ax = plt.subplot(122)
23 # c_counts.plot(kind='bar', ax=ax)
24 # plt.bar(c_counts.index, c_counts.values)
25 # plt.show()
26 '''
27 存在492例盗刷，占总样本的0.17%，
28 存在明显的数据类别不平衡问题，
29 可采用过采样(增加数据)的方法处理该问题
30 '''

1 credit['Time'] = credit['Time'].map(lambda x: divmod(x, 3600)[0])
2 # print(credit['Time'])  # map高级函数：将Time中的每个元素作用于lambda函数

 1 # Class列中值为0的为True，值为1为False，生成的cond0行数不变
 2 # cond0 = credit['Class'] == 0
 3 # Class列中值为0的为False，值为1为True，生成的cond0行数不变
 4 # cond1 = credit['Class'] == 1
 5 # print('cond0 >>>>', len(cond0))
 6 # print('cond1 >>>>', len(cond1))
 7 
 8 # 作图分析
 9 # credit['V1'][cond0].plot(kind='hist', bins=500)
10 # credit['V1'][cond1].plot(kind='hist', bins=50)
11 # plt.show()
12 
13 # 调试查看用
14 # print("credit['V1'] >>>>", credit['V1'])
15 # print('cond0 >>>>', cond0)
16 # print('cond1 >>>>', cond1)
17 
18 # 筛选出存在于V1列中且在cond0中为True的值(284315)
19 # print("credit['V1'][cond0] >>>>", credit['V1'][cond0])
20 # 筛选出存在于V1列中且在cond0中为True的值(492)
21 # print("credit['V1'][cond1] >>>>", credit['V1'][cond1])
22 
23 ''' 作图分析：将每一个特征根据Class的真假进行划分， 图像中两种图形的重合度越大说明该特征对Class的影响越小， 所以需要剔除掉无用的特征 '''
24 # cols = ['V1', 'V2', 'V3', 'V4', 'V5', 'V6', 'V7', 'V8', 'V9', 'V10',
25 #         'V11', 'V12', 'V13', 'V14', 'V15', 'V16', 'V17', 'V18', 'V19', 'V20',
26 #         'V21', 'V22', 'V23', 'V24', 'V25', 'V26', 'V27', 'V28']
27 # 作图：28行，1列，每一行显示一个特征对应的图
28 # plt.figure(figsize=(12, 2800))
29 # for i, col in enumerate(cols):
30 #     ax = plt.subplot(28, 1, i + 1)
31 # density(normed)标准化数据：将过大或过小的数据统一标准化
32 #     credit[col][cond0].plot(kind='hist', bins=500, density=True, ax=ax)
33 #     credit[col][cond1].plot(kind='hist', bins=50, density=True, ax=ax)
34 #
35 #     ax.set_title(col)
36 # plt.show()
37 
38 # 待剔除的列(10列)
39 drops = ['V13', 'V15', 'V20', 'V22', 'V23', 'V24', 'V25', 'V26', 'V27', 'V28']
40 # 删除指定列(axis=1按列，axis=0按行)
41 credit2 = credit.drop(labels=drops, axis=1)
42 print('人眼剔除无用列后 >>>>', credit2.shape)
43 ''' 不同变量在信用卡被盗刷和信用卡正常的不同分布情况， 选择在不同信用卡状态下的分布有明显区别的变量。 
因此剔除变量V13 、V15 、V20 、V22、 V23 、V24 、V25 、V26 、V27 和V28变量 '''

 1 # print('原Amount数据最大值', credit2['Amount'].max())
 2 # print('原Amount数据最小值', credit2['Amount'].min())
 3 # print('原Time数据最大值', credit2['Time'].max())
 4 # print('原Time数据最小值', credit2['Time'].min())
 5 
 6 # 创建标准化对象
 7 standScaler = StandardScaler()
 8 cols = ['Time', 'Amount']
 9 # 标准化数据
10 credit2[cols] = standScaler.fit_transform(credit2[cols])
11 # print('标准化Amount后最大值 >>>>', credit2['Amount'].max())
12 # print('标准化Amount后最小值 >>>>', credit2['Amount'].min())
13 # print('标准化Time后最大值 >>>>', credit2['Time'].max())
14 # print('标准化Time后最小值 >>>>', credit2['Time'].min())

 1 # 创建GBDT对象
 2 # clf = GradientBoostingClassifier()
 3 # 特征训练集：前20列
 4 # X_train = credit2.iloc[:, :-1]
 5 # print('X_train.shape >>>>', X_train.shape)
 6 # cols = X_train.columns
 7 # print('X_train.columns >>>>', X_train.columns)
 8 # 目标值训练集：Class列
 9 # y_train = credit2['Class']  # y_train = credit2.iloc[:,-1]
10 # print('y_train.shape >>>>', y_train.shape)
11 # 训练数据
12 # clf.fit(X_train, y_train)
13 # 得到特征重要性数据
14 # feature_importances_ = clf.feature_importances_
15 # print('feature_importances_ >>>>', feature_importances_)
16 # 从大到小对特征重要性进行排序，并作图分析
17 # argsort()：对数组排序并返回排序后每个元素对应的未排序时自身所在的索引
18 # index = feature_importances_.argsort()[::-1]
19 # print('从大到小排列特征重要性，返回每个元素的原索引 >>>>', index, len(index))
20 
21 # plt.figure(figsize=(12, 9))
22 # 柱状图，第二个参数代表按从大到小排列的特征数据
23 # plt.bar(np.arange(len(index)), feature_importances_[index])
24 # 柱状图x坐标：第二个参数是按特征值从大到小排列后的特征名
25 # plt.xticks(np.arange(len(index)), cols[index])
26 # plt.show()
27 # 根据图像得到要删除的特征列(最小的后9列)
28 drops = ['V7', 'V21', 'V8', 'V5', 'V4', 'V11', 'V19', 'V1', 'Amount']
29 credit3 = credit2.drop(labels=drops, axis=1)
30 print('通过GBDT分析剔除无用列后 >>>>', credit3.shape)
31 # print('credit3.columns >>>>', credit3.columns)

 1 # SMOTE 过采样
 2 X = credit3.iloc[:, :-1]
 3 y = credit3.Class
 4 X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)
 5 X_train,y_train 作为训练数据 训练时，保证样本均衡，将X_train和y_train样本过采样处理 测试时候，可以样本不均衡
 6 # print('未均衡的y训练集分类统计(Class) >>>>', y_train.value_counts())
 7 
 8 smote = SMOTE()
 9 # ndarray
10 X_train_new, y_train_new = smote.fit_sample(X_train, y_train)
11 # print('均衡后的x训练集 >>>>', X_train_new, type(X_train_new))
12 # print('均衡后的y训练集(Class) >>>>', y_train_new, type(y_train_new), len(y_train_new))
13 # y_train_new类型为numpy.ndarray，需转化为pandas.Series类型才可分类统计
14 # print('均衡后的y训练集分类统计(Class) >>>>', Series(y_train_new).value_counts())

 1 # 创建逻辑回归对象
 2 # logistic = LogisticRegression()
 3 # print(logistic)
 4 '''
 5 LogisticRegression(C=1.0, class_weight=None, dual=False, fit_intercept=True,
 6           intercept_scaling=1, max_iter=100, multi_class='ovr', n_jobs=1,
 7           penalty='l2', random_state=None, solver='liblinear', tol=0.0001,
 8           verbose=0, warm_start=False)
 9 '''
10 # 训练均衡后的数据
11 # logistic.fit(X_train_new, y_train_new)
12 # 预测
13 # y_ = logistic.predict(X_test)
14 # print('y_test >>>>', y_test)
15 # print('预测的y_ >>>>', y_)
16 # 交叉表
17 # print('交叉表 >>>>', pd.crosstab(y_test, y_, margins=True))
18 
19 # 混合矩阵
20 # cm = confusion_matrix(y_test, y_)
21 # print('混合矩阵 >>>>', cm, type(cm))
22 # Recall------“正确被检索的正样本item(TP)"占所有"应该检索到的item(TP+FN)"的比例
23 # plot_confusion_matrix(cm, [0, 1], title='Recall:%0.3f' % (cm[1, 1] / (cm[1, 0] + cm[1, 1])))

 1 logistic = LogisticRegression()
 2 clf = GridSearchCV(logistic, param_grid={'tol': [1e-3, 1e-4, 1e-5], 'C': [1, 0.1, 10, 100]}, cv=10, iid=False, n_jobs=1)
 3 print(clf.fit(X_train_new, y_train_new))
 4 # print('best_score_ >>>>', clf.best_score_)
 5 # print('best_params_ >>>>', clf.best_params_)
 6 # print('best_index_ >>>>', clf.best_index_)
 7 # print('best_estimator_ >>>>', clf.best_estimator_)
 8 
 9 # 预测
10 # y3_ = clf.best_estimator_.predict(X_test)
11 # print('y3_预测(best_estimator_) >>>>', confusion_matrix(y_test, y3_))
12 
13 # y2_ = clf.predict(X_test)
14 # print('y2_预测 >>>>', confusion_matrix(y_test, y2_))
15 
16 # cm2 = confusion_matrix(y_test, y2_)
17 
18 # 可视化，对比逻辑斯蒂回归和GridSearchCV结果
19 # plot_confusion_matrix(cm, [0, 1], title='Logistic Recall:%0.3f' % (cm[1, 1] / (cm[1, 0] + cm[1, 1])))
20 # plot_confusion_matrix(cm2, [0, 1], title='GridSearchCV Recall:%0.3f' % (cm2[1, 1] / (cm2[1, 0] + cm2[1, 1])))

1 y_proba = clf.predict_proba(X_test)
2 # 预测被盗刷的概率
3 print(y_proba)