# 导入数据
train = pd.read_excel("BankChurners2.xlsx",sheet_name='training', engine='openpyxl')
test = pd.read_excel("BankChurners2.xlsx",sheet_name='testing', engine='openpyxl')
x_train = train.drop(columns=['id','Attrition_Flag'])
x_test = test.drop(columns=['id','Attrition_Flag'])
y_train = train['Attrition_Flag']
y_test = test['Attrition_Flag']
查看数据¶
Python
x_train.head()
Text Only
##    Customer_Age Gender  ...  Total_Ct_Chng_Q4_Q1 Avg_Utilization_Ratio
## 0            46      M  ...                0.559                 0.289
## 1            58      M  ...                0.778                 0.088
## 2            47      M  ...                1.357                 0.000
## 3            54      M  ...                0.423                 0.000
## 4            44      F  ...                0.667                 0.000
## [5 rows x 19 columns]
Python
y_train.head()
Text Only
## 0    Existing Customer
## 1    Existing Customer
## 2    Existing Customer
## 3    Attrited Customer
## 4    Existing Customer
## Name: Attrition_Flag, dtype: object
特征筛选¶
绘制相关系数热力图，识别特征之间的相关性¶
Python
# 生成相关系数矩阵
corr = x_train.corr()
# 生成上三角形矩阵，在后面可以用 mask 把上三角形的相关系数覆盖住
matrix = np.triu(corr)
# 设置图片大小
fig, ax = plt.subplots(figsize=(20,10))
# 生成热力图。cmap 是颜色，annot 是是否显示相关系数，mask 是上三角形矩阵
sb.heatmap(corr, cmap="Blues", annot=True, mask=matrix)
plt.show()
Python
plt.close()
Python
fig, ax = plt.subplots(figsize=(5,5))
plt.scatter(x_train['Credit_Limit'], x_train['Avg_Open_To_Buy'])
plt.show()
Python
plt.close()
Python
# 由于 Avg_Open_To_Buy 和 Credit_Limit 相关非常高，故删除 Avg_Open_To_Buy
x_train = x_train.drop(columns=['Avg_Open_To_Buy'])
x_test = x_test.drop(columns=['Avg_Open_To_Buy'])
Python
# 离散变量
discrete_features = ['Gender','Dependent_count','Education_Level', 'Marital_Status', 'Income_Category', 'Card_Category', 'Total_Relationship_Count','Months_Inactive_12_mon', 'Contacts_Count_12_mon']
# 连续变量
continuous_features = x_train.columns.drop(discrete_features)
对离散分组变量，计算 Information Value，对特征进行筛选。¶
Python
def cal_IV(feature, label):
    # 计算 WOE
    woe = pd.crosstab(feature, label, normalize='columns')\
    .assign(woe=lambda dfx: np.log(dfx['Attrited Customer'] / dfx['Existing Customer']))
    # 计算 IV
    iv = np.sum(woe['woe']*(woe['Attrited Customer']-woe['Existing Customer']))
    return iv
Python
# 存放各变量的 IV 值
iv_dict = {}
for feature in discrete_features:
    iv_dict[feature] = cal_IV( x_train[feature], y_train)
Python
# 对 IV 值进行排序
iv_dict = dict(sorted(iv_dict.items(), key=lambda item: item[1]))
iv_dict
Text Only
## {'Card_Category': 0.0013000619131457577, 'Marital_Status': 0.0036967489569619365, 'Gender': 0.003925134588828062, 'Income_Category': 0.009186000687735793, 'Education_Level': 0.010361108736073923, 'Dependent_count': 0.011508555906514393, 'Total_Relationship_Count': 0.17613730345552436, 'Months_Inactive_12_mon': 0.375963833865356, 'Contacts_Count_12_mon': inf}
Python
# 由于 Card_Category 的 IV 值非常低，故删除 Card_Category
discrete_features.remove('Card_Category')
x_train = x_train.drop(columns=['Card_Category'])
x_test = x_test.drop(columns=['Card_Category'])
将文本型的特征和标签编码成数值型¶
Python
# 将特征转换为数值型
OrdinalEncoder = preprocessing.OrdinalEncoder()
x_train = OrdinalEncoder.fit_transform(x_train[discrete_features])
x_test = OrdinalEncoder.transform(x_test[discrete_features])
Python
# 将标签转换为数值型
le = preprocessing.LabelEncoder()
y_train = le.fit_transform(y_train)
y_test = le.transform(y_test)
构建决策树¶
基于交叉验证选择最优的树深度¶
Python
# 基于交叉验证选择最优的树深度
depths = np.arange(1, 20)
scores = []
for depth in depths:
    clf = DecisionTreeClassifier(max_depth=depth, random_state=42)
    score = cross_val_score(clf, x_train, y_train, cv=5).mean()
    scores.append(score)
绘制树深度与准确率的关系¶
Python
# 绘制树深度与准确率的关系
plt.plot(depths, scores)
plt.annotate('best depth: {}'.format(depths[np.argmax(scores)]), xy=(depths[np.argmax(scores)], np.max(scores)), xytext=(np.argmax(scores)+1, np.max(scores)-0.02), arrowprops=dict(facecolor='black', shrink=0.05))
plt.xlabel('depth')
plt.ylabel('score')
plt.show()
Python
plt.close()
选择最优的树深度，构建决策树¶
Python
# 选择最优的树深度，构建决策树
clf = DecisionTreeClassifier(max_depth=depths[np.argmax(scores)], random_state=42)
# 训练模型
clf.fit(x_train, y_train)
# 预测标签
Text Only
## DecisionTreeClassifier(max_depth=4, random_state=42)
Python
y_pred = clf.predict(x_test)
# 预测标签的概率
y_pred_proba = clf.predict_proba(x_test)[:,1]
y_pred_proba
Text Only
## array([0.79381919, 0.79381919, 0.79381919, ..., 0.79381919, 0.91443167,
##        0.79381919])
模型评价¶
Accuracy, Presicion, Recall, F1¶
Python
# 计算准确率 Accuracy
accuracy = np.sum(y_pred == y_test) / len(y_test)
# 计算精确率 Presicion
presicion = np.sum(y_pred[y_pred == 1] == y_test[y_pred == 1]) / np.sum(y_pred == 1)
# 计算召回率 Recall
recall = np.sum(y_pred[y_test == 1] == y_test[y_test == 1]) / np.sum(y_test == 1)
# 计算 F1 值
f1 = 2 * presicion * recall / (presicion + recall)
Python
print('Accuracy: {:.2%}\n\
Presicion: {:.2%}\n\
Recall: {:.2%}\n\
F1: {:.2%}'.format(accuracy, presicion, recall, f1))
Text Only
## Accuracy: 86.09%
## Presicion: 85.93%
## Recall: 99.85%
## F1: 92.37%
使用 sklearn 自带的评价指标函数¶
Python
print('Accuracy: {:.2%}\n\
Presicion: {:.2%}\n\
Recall: {:.2%}\n\
F1: {:.2%}'.format(metrics.accuracy_score(y_test, y_pred), metrics.precision_score(y_test, y_pred), metrics.recall_score(y_test, y_pred), metrics.f1_score(y_test, y_pred)))
Text Only
## Accuracy: 86.09%
## Presicion: 85.93%
## Recall: 99.85%
## F1: 92.37%
绘制 ROC 曲线¶
Python
# 绘制 ROC 曲线
fpr, tpr, thresholds = metrics.roc_curve(y_test, y_pred_proba)
plt.plot(fpr, tpr, color='darkorange', lw=2, label='ROC curve')
plt.plot([0, 1], [0, 1], color='navy', lw=2, linestyle='--')
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('ROC Curve')
plt.legend(loc="lower right")
plt.show()
Python
plt.close()
计算 AUC¶
Python
# 计算 AUC
auc = metrics.auc(fpr, tpr)
print('AUC: {:.2%}'.format(auc))
Text Only
## AUC: 71.57%