机器学习实战案例 - 完整项目流程

前置知识：需要先完成机器学习系列前三篇教程的学习

本文重点：完整项目流程实践，建立工程化思维

---

一、项目概述

1.1 问题描述

我们将完成一个客户流失预测项目，这是电信行业的经典问题：

• 目标：预测哪些客户可能流失
• 数据：客户信息、服务订阅、账单等
• 价值：提前识别高风险客户，降低流失率

1.2 项目流程

1. 数据获取与理解
2. 数据探索性分析 (EDA)
3. 数据预处理
4. 特征工程
5. 模型选择与训练
6. 模型评估与调优
7. 模型部署与监控

---

二、数据获取与理解

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
# 设置
plt.rcParams['font.sans-serif'] = ['Arial Unicode MS', 'SimHei']
plt.rcParams['axes.unicode_minus'] = False
pd.set_option('display.max_columns', 30)
# 生成模拟数据
np.random.seed(42)
n_samples = 5000
# 生成客户数据
data = {
    'customer_id': range(1, n_samples + 1),
    'credit_score': np.random.randint(300, 850, n_samples),
    'geography': np.random.choice(['France', 'Germany', 'Spain'], n_samples, p=[0.5, 0.25, 0.25]),
    'gender': np.random.choice(['Male', 'Female'], n_samples),
    'age': np.random.randint(18, 92, n_samples),
    'tenure': np.random.randint(0, 11, n_samples),
    'balance': np.random.exponential(50000, n_samples) * np.random.choice([0, 1], n_samples, p=[0.4, 0.6]),
    'num_of_products': np.random.choice([1, 2, 3, 4], n_samples, p=[0.5, 0.45, 0.04, 0.01]),
    'has_cr_card': np.random.choice([0, 1], n_samples, p=[0.3, 0.7]),
    'is_active_member': np.random.choice([0, 1], n_samples, p=[0.5, 0.5]),
    'estimated_salary': np.random.uniform(10000, 200000, n_samples),
}
df = pd.DataFrame(data)
# 生成目标变量（与特征有一定相关性）
churn_prob = (
    0.1 + 
    0.2 * (df['age'] > 50).astype(int) +
    0.15 * (df['geography'] == 'Germany').astype(int) +
    0.1 * (df['is_active_member'] == 0).astype(int) +
    0.1 * (df['num_of_products'] == 1).astype(int) +
    0.15 * (df['balance'] == 0).astype(int) -
    0.1 * (df['tenure'] > 5).astype(int)
)
churn_prob = np.clip(churn_prob, 0.05, 0.8)
df['exited'] = (np.random.random(n_samples) < churn_prob).astype(int)
print("数据概览:")
print(df.info())
print("\n前5行:")
print(df.head())
print("\n目标变量分布:")
print(df['exited'].value_counts(normalize=True))

---

三、探索性数据分析

# 1. 基础统计
print("=== 基础统计 ===")
print(df.describe())
print("\n类别变量分布:")
for col in ['geography', 'gender']:
    print(f"\n{col}:")
    print(df[col].value_counts())
# 2. 目标变量分布
fig, axes = plt.subplots(1, 2, figsize=(12, 5))
# 计数
df['exited'].value_counts().plot(kind='bar', ax=axes[0])
axes[0].set_title('流失分布')
axes[0].set_xlabel('是否流失')
axes[0].set_ylabel('客户数')
axes[0].set_xticklabels(['未流失', '流失'], rotation=0)
# 比例
df['exited'].value_counts(normalize=True).plot(kind='pie', ax=axes[1], autopct='%1.1f%%')
axes[1].set_title('流失比例')
axes[1].set_ylabel('')
plt.tight_layout()
plt.savefig('churn_distribution.png', dpi=100, bbox_inches='tight')
plt.close()
# 3. 数值变量分布
numeric_cols = ['credit_score', 'age', 'tenure', 'balance', 'estimated_salary']
fig, axes = plt.subplots(2, 3, figsize=(15, 10))
for ax, col in zip(axes.ravel(), numeric_cols):
    df[df['exited'] == 0][col].hist(bins=30, alpha=0.5, label='未流失', ax=ax)
    df[df['exited'] == 1][col].hist(bins=30, alpha=0.5, label='流失', ax=ax)
    ax.set_title(col)
    ax.legend()
axes.ravel()[-1].axis('off')
plt.tight_layout()
plt.savefig('numeric_distribution.png', dpi=100, bbox_inches='tight')
plt.close()
# 4. 类别变量与流失关系
fig, axes = plt.subplots(1, 3, figsize=(15, 5))
for ax, col in zip(axes, ['geography', 'gender', 'num_of_products']):
    pd.crosstab(df[col], df['exited'], normalize='index').plot(kind='bar', ax=ax)
    ax.set_title(f'{col}与流失关系')
    ax.set_xlabel(col)
    ax.set_ylabel('流失率')
    ax.legend(['未流失', '流失'])
    ax.set_xticklabels(ax.get_xticklabels(), rotation=45)
plt.tight_layout()
plt.savefig('categorical_churn.png', dpi=100, bbox_inches='tight')
plt.close()
# 5. 相关性分析
numeric_df = df.select_dtypes(include=[np.number])
plt.figure(figsize=(10, 8))
sns.heatmap(numeric_df.corr(), annot=True, cmap='coolwarm', center=0, fmt='.2f')
plt.title('特征相关性热力图')
plt.tight_layout()
plt.savefig('correlation_heatmap.png', dpi=100, bbox_inches='tight')
plt.close()
# 6. 关键发现
print("\n=== 关键发现 ===")
print("1. 德国客户流失率较高")
print("2. 年龄较大的客户更容易流失")
print("3. 不活跃会员流失风险更高")
print("4. 余额为0的客户流失率较高")

---

四、数据预处理

from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler, LabelEncoder
from sklearn.impute import SimpleImputer
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
# 1. 数据分割
X = df.drop(['customer_id', 'exited'], axis=1)
y = df['exited']
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42, stratify=y
)
print(f"训练集大小: {X_train.shape}")
print(f"测试集大小: {X_test.shape}")
# 2. 定义特征类型
numeric_features = ['credit_score', 'age', 'tenure', 'balance', 'estimated_salary']
categorical_features = ['geography', 'gender']
binary_features = ['has_cr_card', 'is_active_member', 'num_of_products']
# 3. 创建预处理管道
from sklearn.preprocessing import OneHotEncoder
numeric_transformer = Pipeline(steps=[
    ('imputer', SimpleImputer(strategy='median')),
    ('scaler', StandardScaler())
])
categorical_transformer = Pipeline(steps=[
    ('imputer', SimpleImputer(strategy='constant', fill_value='Unknown')),
    ('onehot', OneHotEncoder(handle_unknown='ignore', sparse_output=False))
])
preprocessor = ColumnTransformer(
    transformers=[
        ('num', numeric_transformer, numeric_features),
        ('cat', categorical_transformer, categorical_features),
        ('bin', 'passthrough', binary_features)
    ]
)
# 4. 检查预处理结果
X_train_processed = preprocessor.fit_transform(X_train)
print(f"\n预处理后特征数: {X_train_processed.shape[1]}")

---

五、模型训练与比较

from sklearn.linear_model import LogisticRegression
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
from sklearn.svm import SVC
from sklearn.metrics import (
    accuracy_score, precision_score, recall_score, f1_score,
    roc_auc_score, confusion_matrix, classification_report
)
import time
# 定义模型
models = {
    'Logistic Regression': LogisticRegression(max_iter=1000, random_state=42),
    'Decision Tree': DecisionTreeClassifier(max_depth=10, random_state=42),
    'Random Forest': RandomForestClassifier(n_estimators=100, random_state=42, n_jobs=-1),
    'Gradient Boosting': GradientBoostingClassifier(n_estimators=100, random_state=42),
}
# 训练和评估
results = []
for name, model in models.items():
    # 创建完整管道
    pipe = Pipeline([
        ('preprocessor', preprocessor),
        ('classifier', model)
    ])
    
    # 训练
    start_time = time.time()
    pipe.fit(X_train, y_train)
    train_time = time.time() - start_time
    
    # 预测
    y_pred = pipe.predict(X_test)
    y_prob = pipe.predict_proba(X_test)[:, 1]
    
    # 评估
    metrics = {
        'Model': name,
        'Accuracy': accuracy_score(y_test, y_pred),
        'Precision': precision_score(y_test, y_pred),
        'Recall': recall_score(y_test, y_pred),
        'F1': f1_score(y_test, y_pred),
        'AUC': roc_auc_score(y_test, y_prob),
        'Train Time': train_time
    }
    results.append(metrics)
    
    print(f"\n=== {name} ===")
    print(f"准确率: {metrics['Accuracy']:.4f}")
    print(f"精确率: {metrics['Precision']:.4f}")
    print(f"召回率: {metrics['Recall']:.4f}")
    print(f"F1分数: {metrics['F1']:.4f}")
    print(f"AUC: {metrics['AUC']:.4f}")
# 结果对比
results_df = pd.DataFrame(results)
print("\n=== 模型对比 ===")
print(results_df.set_index('Model'))
# 可视化对比
fig, axes = plt.subplots(1, 2, figsize=(14, 5))
# 性能指标对比
metrics_cols = ['Accuracy', 'Precision', 'Recall', 'F1', 'AUC']
results_df.set_index('Model')[metrics_cols].plot(kind='bar', ax=axes[0])
axes[0].set_title('模型性能对比')
axes[0].set_ylabel('分数')
axes[0].legend(bbox_to_anchor=(1.05, 1), loc='upper left')
axes[0].set_xticklabels(axes[0].get_xticklabels(), rotation=45)
# 训练时间对比
results_df.set_index('Model')['Train Time'].plot(kind='bar', ax=axes[1])
axes[1].set_title('训练时间对比')
axes[1].set_ylabel('时间 (秒)')
axes[1].set_xticklabels(axes[1].get_xticklabels(), rotation=45)
plt.tight_layout()
plt.savefig('model_comparison.png', dpi=100, bbox_inches='tight')
plt.close()

---

六、模型调优

from sklearn.model_selection import GridSearchCV
# 选择最佳模型进行调优 (假设是随机森林)
pipe_rf = Pipeline([
    ('preprocessor', preprocessor),
    ('classifier', RandomForestClassifier(random_state=42, n_jobs=-1))
])
# 参数网格
param_grid = {
    'classifier__n_estimators': [50, 100, 200],
    'classifier__max_depth': [10, 15, 20, None],
    'classifier__min_samples_split': [2, 5, 10],
    'classifier__min_samples_leaf': [1, 2, 4]
}
# 网格搜索
grid_search = GridSearchCV(
    pipe_rf,
    param_grid,
    cv=5,
    scoring='f1',
    n_jobs=-1,
    verbose=1
)
print("=== 开始网格搜索 ===")
grid_search.fit(X_train, y_train)
print(f"\n最优参数: {grid_search.best_params_}")
print(f"最优F1分数: {grid_search.best_score_:.4f}")
# 使用最优模型
best_model = grid_search.best_estimator_
y_pred = best_model.predict(X_test)
y_prob = best_model.predict_proba(X_test)[:, 1]
print("\n=== 最优模型测试集表现 ===")
print(f"准确率: {accuracy_score(y_test, y_pred):.4f}")
print(f"精确率: {precision_score(y_test, y_pred):.4f}")
print(f"召回率: {recall_score(y_test, y_pred):.4f}")
print(f"F1分数: {f1_score(y_test, y_pred):.4f}")
print(f"AUC: {roc_auc_score(y_test, y_prob):.4f}")
# 混淆矩阵
cm = confusion_matrix(y_test, y_pred)
plt.figure(figsize=(8, 6))
sns.heatmap(cm, annot=True, fmt='d', cmap='Blues',
            xticklabels=['未流失', '流失'],
            yticklabels=['未流失', '流失'])
plt.xlabel('预测标签')
plt.ylabel('真实标签')
plt.title('混淆矩阵')
plt.savefig('confusion_matrix_final.png', dpi=100, bbox_inches='tight')
plt.close()
# ROC曲线
from sklearn.metrics import roc_curve
fpr, tpr, thresholds = roc_curve(y_test, y_prob)
plt.figure(figsize=(8, 6))
plt.plot(fpr, tpr, label=f'AUC = {roc_auc_score(y_test, y_prob):.4f}')
plt.plot([0, 1], [0, 1], 'k--', label='随机')
plt.xlabel('假正率')
plt.ylabel('真正率')
plt.title('ROC曲线')
plt.legend()
plt.grid(True, alpha=0.3)
plt.savefig('roc_curve_final.png', dpi=100, bbox_inches='tight')
plt.close()

---

七、特征重要性分析

# 获取特征名称
feature_names = (
    numeric_features + 
    list(best_model.named_steps['preprocessor']
         .named_transformers_['cat']
         .named_steps['onehot']
         .get_feature_names_out(categorical_features).tolist()) +
    binary_features
)
# 获取特征重要性
importances = best_model.named_steps['classifier'].feature_importances_
feature_importance_df = pd.DataFrame({
    'feature': feature_names,
    'importance': importances
}).sort_values('importance', ascending=False)
plt.figure(figsize=(10, 8))
plt.barh(feature_importance_df['feature'][:15], feature_importance_df['importance'][:15])
plt.xlabel('重要性')
plt.ylabel('特征')
plt.title('Top 15 特征重要性')
plt.gca().invert_yaxis()
plt.tight_layout()
plt.savefig('feature_importance_final.png', dpi=100, bbox_inches='tight')
plt.close()
print("\n=== Top 10 重要特征 ===")
print(feature_importance_df.head(10))

---

八、模型部署

import joblib
import json
# 保存模型
joblib.dump(best_model, 'churn_model.joblib')
# 保存特征信息
model_info = {
    'model_name': 'Customer Churn Prediction',
    'version': '1.0',
    'features': {
        'numeric': numeric_features,
        'categorical': categorical_features,
        'binary': binary_features
    },
    'performance': {
        'accuracy': float(accuracy_score(y_test, y_pred)),
        'precision': float(precision_score(y_test, y_pred)),
        'recall': float(recall_score(y_test, y_pred)),
        'f1': float(f1_score(y_test, y_pred)),
        'auc': float(roc_auc_score(y_test, y_prob))
    }
}
with open('model_info.json', 'w') as f:
    json.dump(model_info, f, indent=2)
# 预测函数
def predict_churn(customer_data, model_path='churn_model.joblib'):
    """预测客户流失概率"""
    model = joblib.load(model_path)
    
    if isinstance(customer_data, dict):
        customer_data = pd.DataFrame([customer_data])
    
    prob = model.predict_proba(customer_data)[:, 1][0]
    pred = model.predict(customer_data)[0]
    
    return {
        'churn_probability': float(prob),
        'prediction': int(pred),
        'risk_level': 'High' if prob > 0.7 else ('Medium' if prob > 0.4 else 'Low')
    }
# 测试预测
sample_customer = {
    'credit_score': 600,
    'geography': 'Germany',
    'gender': 'Male',
    'age': 45,
    'tenure': 3,
    'balance': 50000,
    'num_of_products': 1,
    'has_cr_card': 1,
    'is_active_member': 0,
    'estimated_salary': 100000
}
print("\n=== 示例预测 ===")
result = predict_churn(sample_customer)
print(f"客户信息: {sample_customer}")
print(f"预测结果: {result}")

---

九、项目总结

完整流程回顾

"""
项目完成清单:
✓ 1. 问题定义：明确业务目标和成功指标
✓ 2. 数据理解：探索数据特征和分布
✓ 3. 数据准备：处理缺失值、编码、标准化
✓ 4. 特征工程：选择和创建有效特征
✓ 5. 模型选择：比较多个算法
✓ 6. 模型调优：优化超参数
✓ 7. 模型评估：在测试集验证性能
✓ 8. 模型解释：分析特征重要性
✓ 9. 模型部署：保存模型和预测接口
关键发现:
- 年龄、地理位置、活跃状态是流失的关键预测因素
- 德国客户流失率显著高于其他国家
- 不活跃会员流失风险更高
业务建议:
- 针对高风险群体制定挽留策略
- 提升会员活跃度的激励机制
- 关注德国市场的客户满意度
"""

---

参考资源

• Kaggle客户流失预测 - 竞赛案例
• Scikit-learn案例集 - 官方示例
• Machine Learning Mastery - 实战教程
• Google ML Crash Course - 快速入门
• Full Stack Deep Learning - ML工程实践
• Made With ML - 端到端ML教程
• ML项目最佳实践 - GCP最佳实践

---

系列完成：

1. 监督学习算法详解
2. 集成学习方法
3. 无监督学习
4. 模型评估与调优
5. 本文：机器学习实战案例 返回：机器学习基础 最后更新: 2026年4月13日

访问 -- 赞 -- 踩 --

讨论与反馈