本案例數(shù)據(jù)取自kaggle。
這次的案例使用的數(shù)據(jù)做了脫敏處理经宏，可能通過(guò)降維壓縮或是其他的一些方式進(jìn)行了變換處理犀暑。

1、讀入數(shù)據(jù)

先導(dǎo)入常用庫(kù)

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

# 解決坐標(biāo)軸刻度負(fù)號(hào)亂碼
plt.rcParams['axes.unicode_minus'] = False
# 解決中文亂碼問(wèn)題
plt.rcParams['font.sans-serif'] = ['Simhei']

#顯示全部特征
pd.set_option('display.max_columns', None)

%matplotlib inline

讀入數(shù)據(jù)

data=pd.read_csv('creditcard.csv')
data.head()

Class為標(biāo)簽列烁兰，0表示正衬涂鳎客戶，1表示欺詐客戶沪斟。
我們可以看到广辰，V1-V28是經(jīng)過(guò)降維壓縮等轉(zhuǎn)換手段得到的28個(gè)特征，是經(jīng)過(guò)了歸一化的主之，Amount這個(gè)特征表示交易金額择吊，我們是需要將其也進(jìn)行歸一化的。

data.shape

數(shù)據(jù)集維度是(284807, 31)槽奕。

data.info()

數(shù)據(jù)集不存在缺失值且沒(méi)有字符串變量几睛。

data['Class'].value_counts()

標(biāo)簽分布極不平衡。

2史翘、數(shù)據(jù)處理

首先枉长，我們要將Amount標(biāo)準(zhǔn)化冀续，并將不需要用到的Time去掉。

from sklearn.preprocessing import StandardScaler

data['Amount'] = StandardScaler().fit_transform(data['Amount'].reshape(-1,1))
data=data.loc[:,data.columns!='Time']
data.head()

3必峰、樣本不均衡解決方案

由于我們的數(shù)據(jù)極不平衡洪唐，在這里有上采樣（將少數(shù)類增多到和多數(shù)類一樣）和下采樣（減少多數(shù)類使其數(shù)量與少數(shù)類相同）兩種方式。先來(lái)看看下采樣的方式吼蚁。

3.1 下采樣策略

## 下采樣
# 分離特征和標(biāo)簽
X = data.loc[:, data.columns != 'Class']
y = data['Class']

# 計(jì)算少數(shù)類個(gè)數(shù)
number_records_fraud = len(data[data.Class == 1])
# 取得少數(shù)類樣本的索引
fraud_indices = np.array(data[data.Class == 1].index)
# 取得多數(shù)類樣本的索引
normal_indices = data[data.Class == 0].index

# 從多數(shù)類中隨機(jī)選擇與少數(shù)類個(gè)數(shù)相同的樣本數(shù)作為樣本
random_normal_indices = np.random.choice(normal_indices, number_records_fraud, replace = False)
random_normal_indices = np.array(random_normal_indices)

# 合并隨機(jī)取得的0類和全部的1類的索引
under_sample_indices = np.concatenate([fraud_indices,random_normal_indices])

# 根據(jù)索引得到下采樣后的數(shù)據(jù)集
under_sample_data = data.iloc[under_sample_indices,:]

# 分離特征和標(biāo)簽
X_undersample = under_sample_data.ix[:, under_sample_data.columns != 'Class']
y_undersample = under_sample_data.ix[:, under_sample_data.columns == 'Class']

print("正樣本比例(0類): ", len(under_sample_data[under_sample_data.Class == 0])/len(under_sample_data))
print("負(fù)樣本比例(1類): ", len(under_sample_data[under_sample_data.Class == 1])/len(under_sample_data))
print("下采樣后總樣本個(gè)數(shù)為: ", len(under_sample_data))

下采樣后正負(fù)類個(gè)數(shù)相等凭需，樣本總數(shù)減少為984。
劃分測(cè)試集和訓(xùn)練集肝匆。

from sklearn.model_selection import train_test_split

# 切分訓(xùn)練集和測(cè)試集
X_train, X_test, y_train, y_test = train_test_split(X,y,test_size = 0.3, random_state = 0)

print("源數(shù)據(jù)訓(xùn)練集樣本數(shù): ", len(X_train))
print("源數(shù)據(jù)測(cè)試集樣本數(shù): ", len(X_test))
print("源數(shù)據(jù)總樣本數(shù): ", len(X_train)+len(X_test))

# 下采樣
X_train_undersample, X_test_undersample, y_train_undersample, y_test_undersample = train_test_split(X_undersample
                                                                                                   ,y_undersample
                                                                                                   ,test_size = 0.3
                                                                                                   ,random_state = 0)
print('\n')
print("下采樣后訓(xùn)練集樣本數(shù): ", len(X_train_undersample))
print("下采樣后測(cè)試集樣本數(shù): ", len(X_test_undersample))
print("下采樣后樣本總數(shù): ", len(X_train_undersample)+len(X_test_undersample))

3.2 正則化參數(shù)擇優(yōu)

from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import KFold, cross_val_score
from sklearn.metrics import confusion_matrix,recall_score,classification_report 

def printing_Kfold_scores(x_train_data,y_train_data):
    fold = KFold(5,shuffle=False) 

    # 超參數(shù)C范圍
    c_param_range = [0.01,0.1,1,10,100]

    results_table = pd.DataFrame(index = range(len(c_param_range),2), columns = ['C_parameter','Mean recall score'])
    results_table['C_parameter'] = c_param_range

    # the k-fold will give 2 lists: train_indices = indices[0], test_indices = indices[1]
    j = 0
    for c_param in c_param_range:
        print('-------------------------------------------')
        print('超參數(shù)C: ', c_param)
        print('-------------------------------------------')
        print('\n')

        recall_accs = []
        for iteration, indices in enumerate(fold.split(x_train_data)):
            # 建立l1正則化的邏輯回歸模型
            lr = LogisticRegression(C = c_param, penalty = 'l1')
            # 訓(xùn)練模型 
            lr.fit(x_train_data.iloc[indices[0],:],y_train_data.iloc[indices[0],:].values.ravel())

            # 模型預(yù)測(cè)
            y_pred_undersample = lr.predict(x_train_data.iloc[indices[1],:].values)

            # 計(jì)算召回率
            recall_acc = recall_score(y_train_data.iloc[indices[1],:].values,y_pred_undersample)
            recall_accs.append(recall_acc)
            print('Iteration ', iteration,': recall= ', recall_acc)

        # The mean value of those recall scores is the metric we want to save and get hold of.
        results_table.ix[j,'Mean recall score'] = np.mean(recall_accs)
        j += 1
        print('')
        print('Mean recall score ', np.mean(recall_accs))
        print('')

    best_c = results_table.loc[results_table['Mean recall score'].astype('float64').idxmax()]['C_parameter']
    
    # Finally, we can check which C parameter is the best amongst the chosen.
    print('*********************************************************************************')
    print('超參數(shù)C為{}時(shí)粒蜈，模型交叉驗(yàn)證得分最高'.format(best_c))
    print('*********************************************************************************')
    
    return best_c

在0.01、0.1旗国、1枯怖、10、100這五個(gè)不同量級(jí)中選出最優(yōu)的超參數(shù)C(正則化懲罰系數(shù))能曾，評(píng)價(jià)得分標(biāo)準(zhǔn)為recall度硝。

best_c = printing_Kfold_scores(X_train_undersample,y_train_undersample)

3.3 模型評(píng)估

def plot_confusion_matrix(cm, classes,
                          title='Confusion matrix',
                          cmap=plt.cm.Blues):
    """
    This function prints and plots the confusion matrix.
    """
    plt.imshow(cm, interpolation='nearest', cmap=cmap)
    plt.title(title)
    plt.colorbar()
    tick_marks = np.arange(len(classes))
    plt.xticks(tick_marks, classes, rotation=0)
    plt.yticks(tick_marks, classes)

    thresh = cm.max() / 2.
    for i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])):
        plt.text(j, i, cm[i, j],
                 horizontalalignment="center",
                 color="white" if cm[i, j] > thresh else "black")

    plt.tight_layout()
    plt.ylabel('True label')
    plt.xlabel('Predicted label')

import itertools
lr = LogisticRegression(C = best_c, penalty = 'l1')
lr.fit(X_train_undersample,y_train_undersample.values.ravel())
y_pred_undersample = lr.predict(X_test_undersample.values)

# Compute confusion matrix
cnf_matrix = confusion_matrix(y_test_undersample,y_pred_undersample)
np.set_printoptions(precision=2)

print("測(cè)試集Recall: ", cnf_matrix[1,1]/(cnf_matrix[1,0]+cnf_matrix[1,1]))

# Plot non-normalized confusion matrix
class_names = [0,1]
plt.figure()
plot_confusion_matrix(cnf_matrix
                      , classes=class_names
                      , title='Confusion matrix')
plt.show()

測(cè)試集Recall: 0.9387755102040817

上面是模型在下采樣之后的測(cè)試集上的表現(xiàn)，我們?cè)賮?lái)看看在沒(méi)有下采樣的測(cè)試集上的表現(xiàn)寿冕。

lr = LogisticRegression(C = best_c, penalty = 'l1')
lr.fit(X_train_undersample,y_train_undersample.values.ravel())
y_pred = lr.predict(X_test.values)

# Compute confusion matrix
cnf_matrix = confusion_matrix(y_test,y_pred)
np.set_printoptions(precision=2)

print("測(cè)試集Recall: ", cnf_matrix[1,1]/(cnf_matrix[1,0]+cnf_matrix[1,1]))

# Plot non-normalized confusion matrix
class_names = [0,1]
plt.figure()
plot_confusion_matrix(cnf_matrix
                      , classes=class_names
                      , title='Confusion matrix')
plt.show()

測(cè)試集Recall: 0.9183673469387755

可以看到recall略有降低蕊程，但是通過(guò)混淆矩陣可以看到，將正惩粘客戶誤判為欺詐客戶的人數(shù)為9505藻茂，這其實(shí)是很不好的。
我們?cè)倏匆幌氯绻褂梦唇?jīng)過(guò)下采樣的訓(xùn)練集訓(xùn)練出的模型預(yù)測(cè)效果如何玫恳。

lr = LogisticRegression(C = best_c, penalty = 'l1')
lr.fit(X_train,y_train.values.ravel())
y_pred_undersample = lr.predict(X_test.values)

# 計(jì)算混淆矩陣
cnf_matrix = confusion_matrix(y_test,y_pred_undersample)
np.set_printoptions(precision=2)

print("測(cè)試集Recall: ", cnf_matrix[1,1]/(cnf_matrix[1,0]+cnf_matrix[1,1]))

class_names = [0,1]
plt.figure()
plot_confusion_matrix(cnf_matrix
                      , classes=class_names
                      , title='Confusion matrix')
plt.show()

測(cè)試集Recall: 0.5510204081632653

可以看到辨赐，recall值僅有0.55非常低。

3.4 閾值對(duì)結(jié)果的影響

其實(shí)sklearn中的邏輯回歸默認(rèn)概率>0.5判斷為1類纽窟，但我們是可以自己設(shè)定閾值的肖油，不同的閾值會(huì)影響最終的模型評(píng)價(jià)結(jié)果。

lr = LogisticRegression(C = 0.01, penalty = 'l1')
lr.fit(X_train_undersample,y_train_undersample.values.ravel())
y_pred_undersample_proba = lr.predict_proba(X_test_undersample.values)

thresholds = [0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9]

plt.figure(figsize=(10,10))

j = 1
for i in thresholds:
    y_test_predictions_high_recall = y_pred_undersample_proba[:,1] > i
    
    plt.subplot(3,3,j)
    j += 1
    
    # 計(jì)算混淆矩陣
    cnf_matrix = confusion_matrix(y_test_undersample,y_test_predictions_high_recall)
    np.set_printoptions(precision=2)

    print("閾值為{}時(shí)測(cè)試集Recall: {}".format(i,cnf_matrix[1,1]/(cnf_matrix[1,0]+cnf_matrix[1,1])))

    class_names = [0,1]
    plot_confusion_matrix(cnf_matrix
                          , classes=class_names
                          , title='Threshold >= %s'%i) 

3.5 上采樣策略

我們前面看到了臂港，采用下采樣策略，訓(xùn)練出來(lái)的模型對(duì)測(cè)試集預(yù)測(cè)結(jié)果recall大約為0.91视搏，遠(yuǎn)遠(yuǎn)好于不進(jìn)行下采樣的訓(xùn)練模型recall值（僅為0.55）审孽，但是會(huì)將9505名正常客戶誤判為欺詐客戶浑娜。
那么我們?cè)俨捎蒙喜蓸拥姆绞絹?lái)訓(xùn)練模型佑力，看看效果如何。

import pandas as pd
from imblearn.over_sampling import SMOTE
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import confusion_matrix
from sklearn.model_selection import train_test_split

credit_cards=pd.read_csv('creditcard.csv')

columns=credit_cards.columns

features_columns=columns.delete(len(columns)-1)

features=credit_cards[features_columns]
labels=credit_cards['Class']

features_train, features_test, labels_train, labels_test = train_test_split(features, 
                                                                            labels, 
                                                                            test_size=0.2, 
                                                                            random_state=0)
oversampler=SMOTE(random_state=0)
os_features,os_labels=oversampler.fit_sample(features_train,labels_train)

os_features = pd.DataFrame(os_features)
os_labels = pd.DataFrame(os_labels)
best_c = printing_Kfold_scores(os_features,os_labels)

我們來(lái)看下經(jīng)過(guò)SMOTE上采樣后數(shù)據(jù)訓(xùn)練的模型預(yù)測(cè)效果如何筋遭。

lr = LogisticRegression(C = best_c, penalty = 'l1')
lr.fit(os_features,os_labels.values.ravel())
y_pred = lr.predict(features_test.values)

# Compute confusion matrix
cnf_matrix = confusion_matrix(labels_test,y_pred)
np.set_printoptions(precision=2)

print("測(cè)試集Recall: ", cnf_matrix[1,1]/(cnf_matrix[1,0]+cnf_matrix[1,1]))

# Plot non-normalized confusion matrix
class_names = [0,1]
plt.figure()
plot_confusion_matrix(cnf_matrix
                      , classes=class_names
                      , title='Confusion matrix')
plt.show()

測(cè)試集Recall: 0.9108910891089109

可以看到打颤，recall值約為0.91暴拄，與下采樣時(shí)幾乎是一樣，但是僅有536名正潮嘟龋客戶被誤判為了欺詐客戶乖篷，相比于下采樣9505個(gè)誤判值來(lái)說(shuō)，上采樣的方式似乎更優(yōu)透且。

PS：本文代碼參考自《唐宇迪機(jī)器學(xué)習(xí)》