构建用户违约风险预测模型：从数据治理到模型优化的完整流程

在金融行业，预测用户违约风险对于银行和金融机构至关重要。本文将介绍如何构建一个用户违约风险预测模型，从数据治理到模型优化的完整流程。我们将使用 Python 和常用的数据科学库如 Pandas、Scikit-learn 来实现这一过程。

步骤一：数据治理
1.1 读取数据
首先，我们读取 credit.csv 文件，并查看数据的基本信息。

import pandas as pd
import numpy as np

# 读取数据
credit_data = pd.read_csv('/credit.csv')
print("数据概览：")
print(credit_data.head())
print("数据描述：")
print(credit_data.describe())

1.2 填充缺失值
我们需要处理数据中的缺失值。对于数值列，我们使用均值填充；对于非数值列，我们使用前向填充和后向填充。

# 填充数值列的缺失值
numeric_cols = credit_data.select_dtypes(include=[np.number]).columns
credit_data[numeric_cols] = credit_data[numeric_cols].fillna(credit_data[numeric_cols].mean())

# 填充非数值列的缺失值
credit_data = credit_data.fillna(method='ffill').fillna(method='bfill')

1.3 特征工程
将性别用 0 和 1 替换，并对省份列进行 One-Hot 编码。

# 性别编码
credit_data['gender'] = credit_data['gender'].map({'male': 0, 'female': 1})

# 去除列名中的空格
credit_data.columns = credit_data.columns.str.strip()

# One-Hot 编码
province_dummies = pd.get_dummies(credit_data['province'], drop_first=True)
credit_data = pd.concat([credit_data, province_dummies], axis=1)
credit_data.drop(['province'], axis=1, inplace=True)

1.4 处理异常值
使用 Z-score 方法处理数值列中的异常值。

for col in numeric_cols:
    z_score = np.abs((credit_data[col] - credit_data[col].mean()) / credit_data[col].std())
    credit_data = credit_data[z_score < 3]

步骤二：数据分析与挖掘
2.1 数据可视化
绘制反映不同特征条件下的违约情况的图表

import seaborn as sns
import matplotlib.pyplot as plt

# 性别与违约情况
sns.countplot(x='gender', hue='default', data=credit_data)
plt.title('Gender vs Default')
plt.show()

# 年龄与违约情况
sns.histplot(data=credit_data, x='age', hue='default', multiple='stack', kde=True)
plt.title('Age vs Default')
plt.show()

2.2 切分训练集和测试集
按照 8:2 的比例切分训练集和测试集。

from sklearn.model_selection import train_test_split

X = credit_data.drop('default', axis=1)
y = credit_data['default'].astype(int)  # 确保目标变量为整数类型

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

2.3 模型训练与评估
首先使用逻辑回归模型进行训练和评估。

from sklearn.linear_model import LogisticRegression
from sklearn.metrics import classification_report, confusion_matrix

model = LogisticRegression(max_iter=1000)
model.fit(X_train, y_train)

y_pred = model.predict(X_test)
print("分类报告：")
print(classification_report(y_test, y_pred))
print("混淆矩阵：")
print(confusion_matrix(y_test, y_pred))

步骤三：模型优化
3.1 处理数据不平衡
通过上采样方法来平衡数据。

from sklearn.utils import resample

# 分离少数类和多数类
train_majority = credit_data[credit_data.default == 0]
train_minority = credit_data[credit_data.default == 1]

# 上采样少数类
train_minority_upsampled = resample(train_minority, 
                                    replace=True,     # 以放回抽样的方式
                                    n_samples=len(train_majority),    # 使两个类别样本数相同
                                    random_state=123) # 固定随机种子

# 合并上采样的训练集
upsampled = pd.concat([train_majority, train_minority_upsampled])

# 分离特征和目标变量
X_upsampled = upsampled.drop('default', axis=1)
y_upsampled = upsampled['default']

# 重新划分训练集和测试集
X_train, X_test, y_train, y_test = train_test_split(X_upsampled, y_upsampled, test_size=0.2, random_state=42)

3.2 使用随机森林模型进行训练

from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import GridSearchCV

# 使用随机森林训练模型
rf_model = RandomForestClassifier(random_state=42)

# 定义参数网格
param_grid = {
    'n_estimators': [100, 200, 300],
    'max_features': ['sqrt', 'log2'],
    'max_depth': [10, 20, 30, None],
    'criterion': ['gini', 'entropy']
}

# 使用网格搜索进行超参数调优
grid_search = GridSearchCV(estimator=rf_model, param_grid=param_grid, cv=5, scoring='accuracy')
grid_search.fit(X_train, y_train)

# 输出最佳参数
print("最佳参数：")
print(grid_search.best_params_)

# 使用最佳参数训练最终模型
best_rf_model = grid_search.best_estimator_
best_rf_model.fit(X_train, y_train)

# 预测和评估模型
y_pred_rf = best_rf_model.predict(X_test)
print("随机森林分类报告：")
print(classification_report(y_test, y_pred_rf))
print("随机森林混淆矩阵：")
print(confusion_matrix(y_test, y_pred_rf))

完整代码如下：

import pandas as pd
import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split, GridSearchCV
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import classification_report, confusion_matrix
from sklearn.utils import resample
from sklearn.ensemble import RandomForestClassifier

# 1. 读取 credit.csv、查看数据
credit_data = pd.read_csv('./credit.csv')
print("数据概览：")
print(credit_data.head())
print("数据描述：")
print(credit_data.describe())

# 2. 填充表中空值
# 仅填充数值列的缺失值
numeric_cols = credit_data.select_dtypes(include=[np.number]).columns
credit_data[numeric_cols] = credit_data[numeric_cols].fillna(credit_data[numeric_cols].mean())

# 填充非数值列的缺失值
credit_data = credit_data.fillna(method='ffill').fillna(method='bfill')

# 3. 性别用【0、1】替换，省份用 one-hot 编码替换
credit_data['gender'] = credit_data['gender'].map({'male': 0, 'female': 1})
credit_data.columns = credit_data.columns.str.strip()  # 去除列名中的前后空格
province_dummies = pd.get_dummies(credit_data['province'], drop_first=True)
credit_data = pd.concat([credit_data, province_dummies], axis=1)
credit_data.drop(['province'], axis=1, inplace=True)

# 4. 按照合理方法处理异常值（使用 Z-score 方法）
for col in numeric_cols:
    z_score = np.abs((credit_data[col] - credit_data[col].mean()) / credit_data[col].std())
    credit_data = credit_data[z_score < 3]

print("数据处理后的概览：")
print(credit_data.head())
print("数据处理后的描述：")
print(credit_data.describe())

# 5. 画出反映不同特征条件下的违约情况
# 绘制性别与违约情况的关系
sns.countplot(x='gender', hue='default', data=credit_data)
plt.title('Gender vs Default')
plt.show()

# 绘制年龄与违约情况的关系
sns.histplot(data=credit_data, x='age', hue='default', multiple='stack', kde=True)
plt.title('Age vs Default')
plt.show()

# 6. 按照 8:2 切分训练集、测试集
X = credit_data.drop('default', axis=1)
y = credit_data['default'].astype(int)  # 确保目标变量为整数类型
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# 7. 建立用户违约风险模型（逻辑回归模型）
model = LogisticRegression(max_iter=1000)
model.fit(X_train, y_train)
y_pred = model.predict(X_test)
print("分类报告：")
print(classification_report(y_test, y_pred))
print("混淆矩阵：")
print(confusion_matrix(y_test, y_pred))

# 处理数据不平衡（上采样）
train_majority = credit_data[credit_data.default == 0]
train_minority = credit_data[credit_data.default == 1]
train_minority_upsampled = resample(train_minority, 
                                    replace=True,     # 以放回抽样的方式
                                    n_samples=len(train_majority),    # 使两个类别样本数相同
                                    random_state=123) # 固定随机种子

upsampled = pd.concat([train_majority, train_minority_upsampled])
X_upsampled = upsampled.drop('default', axis=1)
y_upsampled = upsampled['default']
X_train, X_test, y_train, y_test = train_test_split(X_upsampled, y_upsampled, test_size=0.2, random_state=42)

# 使用随机森林模型
rf_model = RandomForestClassifier(random_state=42)
param_grid = {
    'n_estimators': [100, 200, 300],
    'max_features': ['sqrt', 'log2'],
    'max_depth': [10, 20, 30, None],
    'criterion': ['gini', 'entropy']
}
grid_search = GridSearchCV(estimator=rf_model, param_grid=param_grid, cv=5, scoring='accuracy')
grid_search.fit(X_train, y_train)

print("最佳参数：")
print(grid_search.best_params_)

best_rf_model = grid_search.best_estimator_
best_rf_model.fit(X_train, y_train)
y_pred_rf = best_rf_model.predict(X_test)
print("随机森林分类报告：")
print(classification_report(y_test, y_pred_rf))
print("随机森林混淆矩阵：")
print(confusion_matrix(y_test, y_pred_rf))

以下是使用线性回归来预测用户的信用额度的示例代码。

# 预测信用额度
import pandas as pd
import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LinearRegression
from sklearn.metrics import mean_squared_error, r2_score

# 1. 读取 credit.csv、查看数据
credit_data = pd.read_csv('./credit.csv')
print("数据概览：")
print(credit_data.head())
print("数据描述：")
print(credit_data.describe())

# 2. 填充表中空值
# 仅填充数值列的缺失值
numeric_cols = credit_data.select_dtypes(include=[np.number]).columns
credit_data[numeric_cols] = credit_data[numeric_cols].fillna(credit_data[numeric_cols].mean())

# 填充非数值列的缺失值
credit_data = credit_data.fillna(method='ffill').fillna(method='bfill')

# 3. 性别用【0、1】替换，省份用 one-hot 编码替换
credit_data['gender'] = credit_data['gender'].map({'male': 0, 'female': 1})
credit_data.columns = credit_data.columns.str.strip()  # 去除列名中的前后空格
province_dummies = pd.get_dummies(credit_data['province'], drop_first=True)
credit_data = pd.concat([credit_data, province_dummies], axis=1)
credit_data.drop(['province'], axis=1, inplace=True)

# 4. 按照合理方法处理异常值（使用 Z-score 方法）
for col in numeric_cols:
    z_score = np.abs((credit_data[col] - credit_data[col].mean()) / credit_data[col].std())
    credit_data = credit_data[z_score < 3]

print("数据处理后的概览：")
print(credit_data.head())
print("数据处理后的描述：")
print(credit_data.describe())

# 5. 选择特征和目标变量
# 我们选择所有特征来预测信用额度
X = credit_data.drop('limit', axis=1)
y = credit_data['limit']

# 6. 按照 8:2 切分训练集、测试集
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# 7. 建立线性回归模型
model = LinearRegression()
model.fit(X_train, y_train)

# 预测和评估模型
y_pred = model.predict(X_test)

# 评估模型性能
print("均方误差 (MSE):", mean_squared_error(y_test, y_pred))
print("决定系数 (R^2):", r2_score(y_test, y_pred))

# 画出真实值和预测值的对比图
plt.scatter(y_test, y_pred)
plt.xlabel("真实值")
plt.ylabel("预测值")
plt.title("真实值 vs 预测值")
plt.show()

精简版

import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.impute import SimpleImputer
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import confusion_matrix, roc_curve, precision_recall_curve, f1_score, roc_auc_score, accuracy_score
import matplotlib.pyplot as plt
import joblib

# 第一步：数据治理
# 1. 读取 credit.csv、查看数据
data = pd.read_csv('/mnt/data/credit.csv')
print("数据概览：")
print(data.head())

# 2. 填充表中空值
imputer = SimpleImputer(strategy='mean')
data.iloc[:, :] = imputer.fit_transform(data)
print("缺失值填充完成")

# 3. 性别用【0、1】替换，省份用 one-hot 编码替换
data['Gender'] = data['Gender'].map({'Male': 1, 'Female': 0})
data = pd.get_dummies(data, columns=['Province'], drop_first=True)
print("性别和省份编码完成")

# 4. 按照合理方法处理异常值
# 例如，去除收入为负数的记录
data = data[data['Income'] >= 0]
print("异常值处理完成")

# 第二步：数据分析与挖掘
# 5. 画出反映不同特征条件下的违约情况
default_by_gender = data.groupby('Gender')['Default'].mean()
default_by_income = data.groupby(pd.cut(data['Income'], bins=5))['Default'].mean()

plt.figure(figsize=(12, 5))
plt.subplot(1, 2, 1)
default_by_gender.plot(kind='bar', title='Default Rate by Gender')
plt.xlabel('Gender')
plt.ylabel('Default Rate')

plt.subplot(1, 2, 2)
default_by_income.plot(kind='bar', title='Default Rate by Income')
plt.xlabel('Income Range')
plt.ylabel('Default Rate')

plt.tight_layout()
plt.savefig('default_rate_by_features.png')
plt.show()

# 6. 按照 8：2 切分训练集、测试集
X = data.drop('Default', axis=1)
y = data['Default']

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
print("数据集切分完成")

# 7. 建立用户违约风险模型
model = RandomForestClassifier(n_estimators=100, random_state=42)
model.fit(X_train, y_train)
print("模型训练完成")

# 保存模型到本地
joblib.dump(model, 'credit_risk_model.joblib')
print("模型已保存")

# 第三步：调参与优化
# 8. 模型参数调优（此处可以使用网格搜索或其他方法进行超参数调整）

# 9. 提交预测模型（包括模型评估）
y_pred = model.predict(X_test)
y_prob = model.predict_proba(X_test)[:, 1]

# 混淆矩阵
conf_matrix = confusion_matrix(y_test, y_pred)
print("混淆矩阵：")
print(conf_matrix)

# PR曲线
precision, recall, _ = precision_recall_curve(y_test, y_prob)
plt.plot(recall, precision)
plt.xlabel('Recall')
plt.ylabel('Precision')
plt.title('PR Curve')
plt.savefig('pr_curve.png')
plt.show()

# ROC曲线
fpr, tpr, _ = roc_curve(y_test, y_prob)
plt.plot(fpr, tpr)
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('ROC Curve')
plt.savefig('roc_curve.png')
plt.show()

# F1指标
f1 = f1_score(y_test, y_pred)
print(f"F1 Score: {f1}")

# AUC指标
auc = roc_auc_score(y_test, y_prob)
print(f"AUC Score: {auc}")

print("模型评估完成")