import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier, BalancedRandomForestClassifier
from sklearn.metrics import classification_report, confusion_matrix, average_precision_score, precision_recall_curve
from imblearn.over_sampling import SMOTE
from imblearn.under_sampling import RandomUnderSampler
from imblearn.pipeline import Pipeline as imbPipeline # 注意：这里使用的是imblearn的Pipeline
from collections import Counter

# 1. 数据加载与初步探索
# 假设我们已经下载了数据集文件 `creditcard_sample.csv`
try:
    df = pd.read_csv('creditcard_sample.csv')
    print(f"数据加载成功！数据集形状：{df.shape}")
except FileNotFoundError:
    print("请先下载数据集文件。")
    # 模拟生成一个小型不平衡数据集用于演示
    np.random.seed(42)
    n_normal, n_fraud = 10000, 100
    df = pd.DataFrame({
        'Time': np.random.randn(n_normal + n_fraud),
        'V1': np.random.randn(n_normal + n_fraud),
        'Amount': np.random.exponential(scale=100, size=n_normal + n_fraud),
        'Class': np.concatenate([np.zeros(n_normal), np.ones(n_fraud)]).astype(int)
    })
    print("已生成模拟数据集。")

print("\n类别分布：")
print(df['Class'].value_counts())
# 输出类似：0: 10000, 1: 100。比例约为100:1，非常不平衡。

# 2. 数据准备
X = df.drop('Class', axis=1)
y = df['Class']

# 标准化特征（对逻辑回归等模型很重要）
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)

# 划分训练集和测试集 - 关键：使用分层抽样(stratify)保持类别比例
X_train, X_test, y_train, y_test = train_test_split(
    X_scaled, y, test_size=0.2, random_state=42, stratify=y
)

print("\n处理后训练集类别分布：", Counter(y_train))
print("处理后测试集类别分布：", Counter(y_test))

# 3. 使用不同方法处理不平衡数据并建模
# 方法A：逻辑回归（基准，无处理）
lr = LogisticRegression(random_state=42, max_iter=1000)
lr.fit(X_train, y_train)
y_pred_lr = lr.predict(X_test)

# 方法B：使用SMOTE过采样
smote = SMOTE(random_state=42)
X_train_smote, y_train_smote = smote.fit_resample(X_train, y_train)
print("\nSMOTE后训练集类别分布：", Counter(y_train_smote))

lr_smote = LogisticRegression(random_state=42, max_iter=1000)
lr_smote.fit(X_train_smote, y_train_smote)
y_pred_lr_smote = lr_smote.predict(X_test)

# 方法C：使用平衡随机森林（集成方法）
brf = BalancedRandomForestClassifier(n_estimators=100, random_state=42, n_jobs=-1)
brf.fit(X_train, y_train)
y_pred_brf = brf.predict(X_test)

# 4. 评估比较
models = {
    'Logistic Regression (原始)': y_pred_lr,
    'Logistic Regression (SMOTE)': y_pred_lr_smote,
    'Balanced Random Forest': y_pred_brf
}

for name, y_pred in models.items():
    print(f"\n{'='*50}")
    print(f"模型：{name}")
    print(classification_report(y_test, y_pred))
    # 特别关注少数类（欺诈，Class=1）的精确率、召回率和F1分数

# 5. 深入分析：平衡随机森林的特征重要性
feature_importances = brf.feature_importances_
feature_names = df.drop('Class', axis=1).columns
feat_imp_df = pd.DataFrame({'Feature': feature_names, 'Importance': feature_importances})
feat_imp_df = feat_imp_df.sort_values(by='Importance', ascending=False).head(10)

plt.figure(figsize=(10, 6))
sns.barplot(x='Importance', y='Feature', data=feat_imp_df)
plt.title('平衡随机森林 - 前10重要特征')
plt.tight_layout()
plt.show()