build_lightgbm.py

import pandas as pd
import numpy as np
import lightgbm as lgb
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report, roc_auc_score, confusion_matrix
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings('ignore')

print("=" * 60)
print("LightGBM 模型：预测客户未来 3 个月资产提升至 100 万+的概率")
print("=" * 60)

# 1. 读取数据
print("\n[1] 读取数据...")
df_base = pd.read_csv('customer/customer_base.csv')
df_behavior = pd.read_csv('customer/customer_behavior_assets.csv')

# 排序
df_behavior = df_behavior.sort_values(['customer_id', 'stat_month']).reset_index(drop=True)

print(f"    客户基础数据：{len(df_base)} 条")
print(f"    客户行为数据：{len(df_behavior)} 条")

# 2. 构造时序特征函数
def create_time_series_features(group, base_month, lookback=6):
    """
    为每个客户在特定基线月份构造时序特征
    group: 某个客户的历史数据
    base_month: 基线月份
    lookback: 往回看几个月
    """
    months_list = sorted(group['stat_month'].unique())
    if base_month not in months_list:
        return None

    base_idx = months_list.index(base_month)
    if base_idx < lookback - 1:
        return None

    feature_months = months_list[base_idx - lookback + 1: base_idx + 1]
    feature_data = group[group['stat_month'].isin(feature_months)].copy()

    if len(feature_data) < lookback:
        return None

    future_months = []
    base_year, base_m = int(base_month.split('-')[0]), int(base_month.split('-')[1])
    for i in range(1, 4):
        m = base_m + i
        y = base_year
        while m > 12:
            m -= 12
            y += 1
        future_months.append(f"{y}-{m:02d}")

    future_data = group[group['stat_month'].isin(future_months)]

    features = {}

    base_data = feature_data[feature_data['stat_month'] == base_month].iloc[0]
    features['total_assets'] = base_data['total_assets']
    features['deposit_balance'] = base_data['deposit_balance']
    features['financial_balance'] = base_data['financial_balance']
    features['fund_balance'] = base_data['fund_balance']
    features['insurance_balance'] = base_data['insurance_balance']
    features['product_count'] = base_data['product_count']
    features['app_login_count'] = base_data['app_login_count']
    features['app_financial_view_time'] = base_data['app_financial_view_time']
    features['credit_card_monthly_expense'] = base_data['credit_card_monthly_expense']
    features['investment_monthly_count'] = base_data['investment_monthly_count']
    features['financial_repurchase_count'] = base_data['financial_repurchase_count']

    assets = feature_data.sort_values('stat_month')['total_assets'].values

    if len(assets) >= 2:
        features['asset_growth_rate_1m'] = (assets[-1] - assets[-2]) / (assets[-2] + 1)
    if len(assets) >= 4:
        features['asset_growth_rate_3m'] = (assets[-1] - assets[-4]) / (assets[-4] + 1)
    if len(assets) >= 6:
        features['asset_growth_rate_6m'] = (assets[-1] - assets[-6]) / (assets[-6] + 1)

    features['asset_mean_6m'] = np.mean(assets)
    features['asset_std_6m'] = np.std(assets)
    features['asset_cv_6m'] = np.std(assets) / (np.mean(assets) + 1)
    features['current_vs_avg'] = (assets[-1] - np.mean(assets)) / (np.mean(assets) + 1)

    if len(assets) >= 2:
        x = np.arange(len(assets))
        slope, intercept = np.polyfit(x, assets, 1)
        features['trend_slope'] = slope
        y_pred = slope * x + intercept
        ss_res = np.sum((assets - y_pred) ** 2)
        ss_tot = np.sum((assets - np.mean(assets)) ** 2)
        features['trend_r_squared'] = 1 - (ss_res / (ss_tot + 1))

    if len(assets) >= 6:
        recent_3_mean = np.mean(assets[-3:])
        earlier_3_mean = np.mean(assets[-6:-3])
        features['asset_momentum'] = (recent_3_mean - earlier_3_mean) / (earlier_3_mean + 1)

    if len(future_data) > 0:
        max_assets_future = future_data['total_assets'].max()
        features['target'] = 1 if max_assets_future >= 1000000 else 0
    else:
        return None

    features['base_month'] = base_month
    features['customer_id'] = group['customer_id'].iloc[0]

    return features

# 3. 构造训练数据
print("\n[2] 构造时序特征和目标变量...")

base_months = ['2024-07', '2024-08', '2024-09', '2024-10', '2024-11',
               '2024-12', '2025-01', '2025-02', '2025-03']

all_samples = []

for customer_id, group in df_behavior.groupby('customer_id'):
    for base_month in base_months:
        result = create_time_series_features(group, base_month, lookback=6)
        if result is not None:
            all_samples.append(result)

df_samples = pd.DataFrame(all_samples)
print(f"    构造样本数：{len(df_samples)}")

# 4. 合并静态特征
print("\n[3] 合并静态特征...")

static_features = ['age', 'monthly_income', 'gender', 'occupation_type',
                   'lifecycle_stage', 'marriage_status', 'city_level']

df_static = df_base[['customer_id'] + static_features].copy()

df_train = df_samples.merge(df_static, on='customer_id', how='left')
print(f"    合并后样本数：{len(df_train)}")

# 5. 数据预处理
print("\n[4] 数据预处理...")

target_col = 'target'
feature_cols = [col for col in df_train.columns if col not in ['target', 'customer_id', 'base_month']]

X = df_train[feature_cols].copy()
y = df_train[target_col].copy()

from sklearn.preprocessing import LabelEncoder
categorical_cols = ['gender', 'occupation_type', 'lifecycle_stage', 'marriage_status', 'city_level']

for col in categorical_cols:
    le = LabelEncoder()
    X[col] = le.fit_transform(X[col].astype(str))

X = X.fillna(0)

print(f"    特征数：{len(feature_cols)}")
print(f"    正样本数：{y.sum()} ({y.mean()*100:.2f}%)")
print(f"    负样本数：{len(y) - y.sum()} ({(1-y.mean())*100:.2f}%)")

# 6. 划分训练集和测试集
print("\n[5] 划分数据集...")

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.3, random_state=42, stratify=y
)

print(f"    训练集：{len(X_train)} 条")
print(f"    测试集：{len(X_test)} 条")

# 7. 训练 LightGBM 模型
print("\n[6] 训练 LightGBM 模型...")

model = lgb.LGBMClassifier(
    n_estimators=100,
    max_depth=6,
    learning_rate=0.1,
    random_state=42,
    class_weight='balanced',
    verbosity=-1
)

model.fit(X_train, y_train)

# 8. 模型评估
print("\n[7] 模型评估...")

y_pred = model.predict(X_test)
y_pred_proba = model.predict_proba(X_test)[:, 1]

print("\n分类报告:")
print(classification_report(y_test, y_pred, target_names=['未达 100 万', '达到 100 万']))

auc_score = roc_auc_score(y_test, y_pred_proba)
print(f"AUC 分数：{auc_score:.4f}")

cm = confusion_matrix(y_test, y_pred)
print("\n混淆矩阵:")
print(cm)

# 9. 特征重要性
print("\n[8] 特征重要性:")
print("-" * 60)

feature_importance = pd.DataFrame({
    'feature': feature_cols,
    'importance': model.feature_importances_
})

feature_importance = feature_importance.sort_values('importance', ascending=False)

print("\n特征重要性（按重要性排序 Top 20）:")
print(feature_importance.head(20)[['feature', 'importance']].to_string(index=False))

feature_importance.to_csv('lgbm_feature_importance.csv', index=False)
print("\n特征重要性已保存到 lgbm_feature_importance.csv")

# 10. 特征重要性可视化
print("\n[9] 生成特征重要性可视化...")

plt.rcParams['font.sans-serif'] = ['SimHei', 'Microsoft YaHei']
plt.rcParams['axes.unicode_minus'] = False

fig, ax = plt.subplots(figsize=(14, 12))

top_features = feature_importance.head(20)

bars = ax.barh(range(len(top_features)), top_features['importance'], color='#667eea')

ax.set_yticks(range(len(top_features)))
ax.set_yticklabels(top_features['feature'], fontsize=11)

for i, (bar, imp) in enumerate(zip(bars, top_features['importance'])):
    width = bar.get_width()
    ax.text(width + 0.1,
            bar.get_y() + bar.get_height()/2,
            f'{imp:.1f}',
            va='center', ha='left',
            fontsize=10)

ax.set_xlabel('特征重要性 (Feature Importance)', fontsize=12)
ax.set_title('LightGBM 特征重要性 Top 20\n(按重要性排序)', fontsize=14, fontweight='bold')

plt.tight_layout()
plt.savefig('lgbm_feature_importance.png', dpi=150, bbox_inches='tight')
print("    特征重要性可视化已保存到 lgbm_feature_importance.png")

# 11. 显示关键发现
print("\n[10] 关键发现:")
print("-" * 60)

top_5_features = feature_importance.head(5)[['feature', 'importance']]

print("\n最重要的特征（Top 5）:")
for idx, row in top_5_features.iterrows():
    print(f"    {row['feature']}: {row['importance']:.1f}")

print("\n" + "=" * 60)
print("LightGBM 模型训练完成！")
print("=" * 60)
print("\n生成的文件:")
print("  1. lgbm_feature_importance.csv - 特征重要性数据")
print("  2. lgbm_feature_importance.png - 特征重要性图")