可解释 AI：SHAP 与 LIME¶

预测值 = 基准值（base value） + Σ SHAP值(特征i)

import numpy as np
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split

# ===== 1. 准备：训练一个预测 2 型糖尿病的随机森林 =====
np.random.seed(42)

# 模拟基因表达 + 临床特征数据
n_samples = 500
feature_names = [
    'gene_PPARG', 'gene_TCF7L2', 'gene_KCNJ11',  # 已知糖尿病相关基因
    'BMI', 'age', 'fasting_glucose', 'HbA1c',     # 临床指标
    'gene_FTO', 'gene_SLC30A8', 'gene_IGF2BP2'   # 更多基因
]

X, y = make_classification(
    n_samples=n_samples,
    n_features=len(feature_names),
    n_informative=6,   # 6 个真正有信息量的特征
    n_redundant=2,
    random_state=42
)

X_df = pd.DataFrame(X, columns=feature_names)
X_train, X_test, y_train, y_test = train_test_split(X_df, y, test_size=0.2, random_state=42)

# 训练随机森林
rf_model = RandomForestClassifier(n_estimators=100, random_state=42)
rf_model.fit(X_train, y_train)
print(f'模型准确率: {rf_model.score(X_test, y_test):.3f}')

# ===== 2. SHAP 分析 =====
# pip install shap
import shap

# TreeExplainer：专为树模型优化，速度极快
explainer = shap.TreeExplainer(rf_model)

# 计算测试集的 SHAP 值
# shap_values 形状: [类别数, 样本数, 特征数]
shap_values = explainer.shap_values(X_test)

# 对二分类问题，取类别1（患病）的 SHAP 值
shap_values_class1 = shap_values[1]  # [样本数, 特征数]

print(f'\nSHAP 值形状: {shap_values_class1.shape}')

# --- 全局特征重要性 ---
# 每个特征的平均绝对 SHAP 值 = 全局重要性
mean_shap = np.abs(shap_values_class1).mean(axis=0)
feature_importance_df = pd.DataFrame({
    'feature': feature_names,
    'mean_abs_shap': mean_shap
}).sort_values('mean_abs_shap', ascending=False)
print('\n全局特征重要性（SHAP）:')
print(feature_importance_df.to_string(index=False))

# --- 单样本解释 ---
sample_idx = 0  # 第一个测试样本
sample_shap = shap_values_class1[sample_idx]  # 该样本每个特征的 SHAP 值
base_value = explainer.expected_value[1]       # 基准预测值（所有样本的平均预测）

print(f'\n样本 {sample_idx} 的预测解释:')
print(f'基准值 (base value): {base_value:.3f}')
for feat, shap_val in zip(feature_names, sample_shap):
    direction = '↑' if shap_val > 0 else '↓'  # 正值增加患病概率，负值降低
    print(f'  {feat}: SHAP={shap_val:+.3f} {direction}')
print(f'最终预测值: {base_value + sample_shap.sum():.3f}')

# 可视化（需要 matplotlib）
try:
    import matplotlib.pyplot as plt

    # 蜂巢图：展示所有样本所有特征的 SHAP 值分布
    shap.summary_plot(shap_values_class1, X_test, feature_names=feature_names, show=False)
    plt.tight_layout()
    plt.savefig('/tmp/shap_summary.png', dpi=100, bbox_inches='tight')
    print('\nSHAP summary plot 已保存到 /tmp/shap_summary.png')
    plt.close()
except Exception as e:
    print(f'可视化跳过: {e}')

# ===== 3. LIME 分析 =====
# pip install lime
try:
    import lime
    import lime.lime_tabular

    # 创建 LIME 解释器
    lime_explainer = lime.lime_tabular.LimeTabularExplainer(
        X_train.values,              # 训练数据（用于估计特征分布）
        feature_names=feature_names,
        class_names=['正常', '糖尿病'],
        mode='classification'
    )

    # 解释单个样本
    sample = X_test.iloc[0].values  # 取第一个测试样本

    explanation = lime_explainer.explain_instance(
        sample,
        rf_model.predict_proba,      # 模型预测函数
        num_features=5               # 展示最重要的 5 个特征
    )

    print('\nLIME 解释（前5个重要特征）:')
    for feat, weight in explanation.as_list():
        print(f'  {feat}: {weight:+.4f}')
except ImportError:
    print("需要安装 lime: pip install lime")